CN113962364B - A Multi-factor Power Load Forecasting Method Based on Deep Learning - Google Patents

Abstract

The invention discloses a multi-factor power load forecasting method based on deep learning, which comprises the steps of firstly completing the acquisition and storage of data, including power load data and environmental influence data; preprocessing and standardizing data based on abnormal data detection, autoregressive interpolation and sequence data normalization of a k-proximity algorithm and an improved DBSCAN algorithm; then, an improved CNN-LSTM electrical load prediction model is provided, and a CNN characteristic extraction module is adopted to learn local characteristics of input data; then inputting the data into an LSTM sequence learning model, and extracting sequence characteristic information of input data; meanwhile, introducing a self-attribute mechanism into the LSTM for learning the characteristics of the LSTM hidden layer, and realizing the extraction of key characteristics by distributing different attention weights so as to improve the final prediction precision; and finally, predicting the electrical load. The invention can promote the digital upgrading of the power grid, meet the individual requirements of users, and realize the association analysis of industry, the dispatching of power generation, the prediction of power utilization trend, the guidance of repeated work and production and the like.

Description

A multi-factor electricity load forecasting method based on deep learning

Technical Field

The present invention relates to the intersection of artificial intelligence and power load forecasting, and more specifically to a multi-factor user power load forecasting method based on deep learning.

Background Art

In recent years, with the rapid and high-quality development of the national economy, people's living standards have continued to improve, and the demand for electricity has also continued to increase. The country has also accelerated the deployment of power projects to ensure sufficient power supply. However, due to the backwardness of the power generation link of related power generation projects, decision makers in the power sector cannot accurately grasp the changes in power grid load, which will lead to problems such as decision-making errors, and cannot maintain a reasonable power supply and demand relationship, resulting in a large amount of power resources. Predicting user power load has the following significance: (1) Mining user power consumption patterns, analyzing user power consumption information, mining user power consumption patterns, and combining factors such as region and time to further explore user power consumption patterns and the reasons for differences. The prediction results of power grid load can help the power department to dispatch power resources and provide personalized services to different users in different regions; (2) The geographical scale guides power transmission. The prediction of power grid load on the geographical scale can be combined with the differences in power generation between regions to facilitate the power department to reasonably dispatch and transmit power resources; (3) Assist in the realization of a visual interactive system, build a prediction and visualization analysis model based on power data for the power load prediction results, and use a prediction method with good analysis and comparison effects to realize the power grid load prediction system, so as to vividly display the predicted power load of different regions, different time periods, and different types of power users. A series of influencing factors of power load can also be reflected in the visualization system.

The power grid load forecasting system takes into account factors such as users, regions, time, environment, and climate, and realizes the prediction of power load at a certain time or period in the future by building a training model. Previously, some methods simulated the logistic growth curve model, multivariate linear regression analysis model, gray prediction model, and neural network model to establish a prediction model, and predicted the power supply load and power consumption in Harbin. A method with higher accuracy was determined through comparative analysis. Some methods are based on BP neural networks, considering the influence of multiple factors such as temperature, and obtained output results with higher accuracy. Some researchers proposed a Lasso-PCA data reduction and feature extraction model to reduce the amount of calculation and parameters of the model and improve the execution efficiency of the model. They also innovatively used an improved adaptive genetic algorithm to optimize the BP neural network training process. The analysis and demonstration showed that this method has higher prediction accuracy for power load. In view of the defects of single-scale research, a recurrent neural network prediction model based on the back propagation algorithm of time series decomposition can comprehensively establish a prediction model based on the influence of multiple factors to predict the power consumption of users in future periods. Some researchers have also proposed an electricity consumption forecasting model based on the water wave optimization algorithm to improve the radial basis function neural network, solving the problem of inaccurate electricity load forecasting based on user levels. By optimizing the hidden layer center of the RBF neural network and the optimization of the expansion constant parameters, it is further proved that this method has higher accuracy.

With the development of smart grids and the increasing demand of the power industry, the importance of power load forecasting has become increasingly apparent, and the requirements for power load forecasting accuracy have become increasingly higher. Summarizing the current similar research and technical findings: Most of the existing power grid load forecasting methods either tend to be traditional algorithms, such as regression analysis and time series methods, which have defects and cannot consider the influence of complex factors such as meteorological data, and the prediction results are very different from the true value; or use machine learning methods that only face a single data source, such as current power load and historical power load. Although the short-term prediction effect is better than that of traditional methods, the user's power load is often greatly affected by environmental factors, such as weather, temperature, humidity, holidays, etc., which makes the previous method poor in prediction accuracy and large deviation in prediction results. Therefore, exploring new multi-factor power load forecasting based on deep learning has become a current research hotspot.

Summary of the invention

In view of the above problems, the purpose of the present invention is to provide a multi-factor user power load prediction method based on deep learning, so as to achieve the purpose of accurately predicting power load under different user categories, regions, and different environmental factors. The technical solution is as follows:

A multi-factor power load forecasting method based on deep learning, comprising the following steps:

Step 1: Obtain electricity load data including different regions, years and electricity consumption categories, as well as environmental impact data including temperature, humidity and wind power, and store them in a sqlite database;

Step 2: Clean and preprocess the power load data and environmental impact data, including: abnormal data detection based on the k-nearest neighbor algorithm and the improved DBSCAN algorithm, autoregressive interpolation, and sequence data normalization;

Step 3: Introduce the self-attention mechanism and reconstruct the features of the hidden layer of LSTM to achieve end-to-end deployment, thereby building a power load data prediction neural network model based on the improved CNN-LSMT, and use the power load data and environmental impact data processed in step 2 as training sets and test sets;

Step 4: Power load forecasting is performed using the improved CNN-LSMT power load data prediction neural network model.

Furthermore, the electricity load data includes user number, region, urban network/rural network, user classification, electricity consumption category, customer category, voltage level, industry category, contract capacity, operating capacity, first power supply date, frozen electricity date and 15-minute frozen electricity; the environmental impact data includes current date, region, maximum temperature, minimum temperature, average humidity, average wind speed, weather type, current temperature for 15 minutes, and whether it is a holiday; and the time of the environmental impact data is synchronized with the time of the electricity load data.

Furthermore, the abnormal data detection based on the k-nearest neighbor algorithm and the improved DBSCAN algorithm in step 2 is specifically as follows:

Step 2.1: Define the average distance between a sample and its neighboring samples as the abnormality score of the sample, use the improved KNN algorithm to obtain the abnormality score of the power load data at each moment, and take the total distance from each cluster to its k neighbors as the final abnormality score;

The k neighboring sets N _k ( _ci ) of the power load data _ci at a certain time i are expressed as:

表示c_i的k个邻近点之一，d_k(c_i)为c_i的k个邻近的平均距离，

表示c_i与其邻近点

的距离；In the formula,

represents one of the k neighbors of _ci , d _k ( _ci ) is the average distance of the k neighbors of _ci ,

represents _ci and its neighboring points

distance;

Then the abnormal score of the electric load data _ci is expressed as:

Where N _k ( _ci ) represents the set of k neighboring points of _ci

Finally, the first m clusters in the anomaly score sorting list are output as the outliers of the power load data;

Step 2.2: Using the improved clustering-based anomaly detection algorithm DBSCAN, first use local parameters to achieve density clustering of small sample data, then iteratively cluster the local clustering results to achieve the final global clustering results, and mark the points that do not belong to any cluster as outliers; specifically, including:

Step a) Parameter update:

Set the size of the cluster sliding window M, calculate the average distance difference of the power load data in the window, set the first k neighboring power loads as MinPst, and set the Euclidean distance between the power load data as Eps;

A weight is set for each power load data to reduce the impact on the final clustering result. The calculation formula of the weight w( _ci , _cj ) is as follows:

Wherein, Cov( _ci , _cj ) is the covariance of the power load data _ci at time i and the power load data cj at time _j , Var( _ci ) is the variance of the power load data _ci at time i, Var( _cj ) is the variance of the power load data cj at time _j ; r( _ci , _cj ) is expressed as the correlation coefficient between _ci and _cj , the smaller the value, the greater the correlation;

Step b) Obtaining the power load data anomaly detection result by analyzing the clustering results:

In the clustering process, the first marked outlier is set as a candidate outlier, and the anomaly score is set plus 1. In the cyclic iterative clustering process, the next candidate outlier is entered and the anomaly score is updated; if the anomaly score S is equal to the number of clusters C, it is marked as an outlier.

Furthermore, the autoregressive interpolation in step 2 has:

The Lagrange interpolation method is used to complete the missing power load data values, so that the n-1 polynomial y= _a0 + _a1x + _a2x2 +…+a _n- 1xn ^-1 passes through the coordinates of n points ( _x1 , ^y1 ), ( _x2 , _y2 ), ( _x3 , _y3 ),… _, ( _xn , _yn ). Then the function expression of Lagrange interpolation is expressed as:

Among them, _xi and _xj represent the i-th and j-th moments of the power user respectively, _yi represents the power load of the power user at the i-th moment; and n represents the total number of moments of power load.

Furthermore, the processing formula for normalizing the sequence data in step 2 is:

Among them, X＝{X ₁ ,X ₂ ,X ₃ …,X _N } represents the same category of electric load data or environmental impact data; its normalized value is Y＝{Y ₁ ,Y ₂ ,Y ₃ …,Y _N }, l∈[1,N], and N is the total number of electric load data or environmental impact data of a certain category.

Furthermore, the step 3 specifically includes:

Step 3.1: Loading and preparing network model data

Load the dataset and divide it into training set, validation set and test set, then construct Dataloader as data reader for the training set and test set respectively. When training the model, calculate the data of each batch. Dataloader loads the data into memory according to the batch size and shuffles the order of each batch of data to improve the robustness of model training.

Step 3.2: Construct a neural network model for power load data prediction based on improved CNN-LSMT

The neural network model includes three parts: CNN feature learning module, LSTM sequence learning module and self-attention mechanism module:

1) The CNN feature learning module includes three one-dimensional convolutional layers, wherein a MaxPooling layer and a ReLu layer are added between two consecutive convolutional layers; the features of the standardized power load data and the environmental impact data are learned through the convolution operation as the feature map output by the convolutional layer; the MaxPooling layer is added to alleviate the limitation of the invariance of the generated feature map, and the activation function ReLu is used to enhance the model's ability to learn complex structures;

2) The LSTM sequence learning module includes three LSTM layers, each layer contains twenty neurons; the first two LSTM layers output the complete sequence of hidden states, and the last LSTM layer outputs the last time step of the hidden state;

The input of the LSTM layer is the influencing parameter xt at time _t and the predicted value _ht-1 at the previous time. The predicted value _ht at time t is obtained through the prediction function F. Its function expression is as follows:

h _t =F(x _t ,h _t-1 )

3) A self-attention mechanism module is added between each of the three LSTM layers, and the self-attention mechanism module assigns weights to the features of the hidden layer extracted by the LSTM layer, thereby mining more discriminative features of the power load data and the environmental impact data;

The output sequence of the hidden layer of the penultimate layer of the LSTM network at time t is assigned the self-attention weight w _tl , which is expressed as:

Among them, L _h is the length of the sequence output by the LSTM hidden layer, l represents the sequence number of the LSTM hidden layer output sequence, and s _tl represents the direct similarity between the sequence l of the LSTM hidden layer at time t and other sequences;

Multiply the feature h _tl of sequence l by its corresponding weight w _tl to form a new feature sequence h _t ' _l and input it into the next LSTM;

Step 3.3: Define the optimizer and set model training parameters

Use mean absolute error as the loss function to monitor the validation loss, set the adaptive optimizer Adam, and adaptively update the learning rate during the training process; and set the number of training iterations epoch, the initial loss function, and the batch size.

Furthermore, the prediction of power load in step 4 includes: ultra-short-term power load prediction, short-term power load prediction and medium- and long-term power load prediction.

The beneficial effects of the present invention are:

1) The multi-factor power load forecasting method based on deep learning proposed in this invention makes up for the shortcomings of the current situation that it only focuses on a single data source, cannot consider the influence of complex factors such as meteorological data, and the prediction results are very different from the true value. By introducing external environmental factors such as temperature, humidity, wind power, etc. as the most influential factors of power load and combining historical power load data as input features, the current power load can be predicted, and the prediction accuracy can be effectively improved after considering environmental factors.

2) The present invention proposes a multi-factor electricity load forecasting method based on deep learning, and proposes a method for detecting abnormal electricity load data by combining KNN and improved DBSCAN, which can effectively mine the discrete values of electricity load data values, and reduce the impact of abnormal values on the prediction model by restoring missing or abnormal electricity load data.

3) The multi-factor power load forecasting method based on deep learning proposed in the present invention proposes an improved CNN-LSTM power load forecasting model. First, the CNN feature extraction module is used to learn the local features of the input data; then it is input into the LSTM sequence learning model to extract the sequence feature information of the input data; at the same time, the self-attention mechanism is introduced into the LSTM to learn the features of the LSTM hidden layer, and the key special extraction is achieved by assigning different attention weights to improve the final prediction accuracy; finally, the power load is predicted through the Dropout layer and the FC layer. In addition, the improved CNN-LSTM prediction model can not only obtain a higher prediction accuracy, but also simplify the neural network construction process, and can achieve end-to-end deployment, thereby promoting the rapid use of research results.

4) The multi-factor electricity load forecasting method based on deep learning proposed in the present invention realizes the prediction of ultra-short-term, short-term and medium- and long-term electricity loads of different categories of users, and visualizes the prediction results, and assists in the analysis of electricity consumption before and after the epidemic, electricity consumption distribution analysis and electricity user portraits. The present invention can promote the digital upgrade of the power grid, meet the personalized needs of users, realize industry correlation analysis, dispatch power generation, predict electricity consumption trends and guide resumption of work and production. The present invention can be used as an independent system or as a component embedded in the original system of the power department, which can save resources to a great extent and improve work efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of the overall process of the present invention.

FIG. 2 is a schematic diagram of a data processing flow of the present invention.

FIG3 is a schematic diagram of the structure of the electricity load prediction model based on the improved CNN-LSTM proposed in the present invention.

Figure 4 is a line graph comparing the ultra-short-term commercial electricity load forecast value and the actual value for a certain region according to the present invention; (a) is a comparison between the forecast value and the actual value of the May 1 commercial electricity load in the region in 2019; (b) is a comparison of the May 1 commercial electricity load in the region in 2018/2019/2020.

FIG5 is a line graph showing a comparison between the predicted value of short-term commercial electricity load in a certain region and the actual value according to the present invention.

FIG6 is a line graph comparing the medium- and long-term electricity load forecast value and the actual value of commercial electricity consumption in a certain area according to the present invention; (a) is the weekly electricity load forecast for the area; (b) is the monthly electricity load forecast for the area.

FIG. 7 is a schematic diagram of an application scheme of the present invention.

DETAILED DESCRIPTION

The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention introduces external environmental factors such as temperature, humidity, wind power, etc. as the most influential factors of power load, and uses abnormal data detection, autoregressive interpolation and sequence data normalization based on k-nearest neighbor (KNN) algorithm and improved DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm to preprocess and standardize the data; then an improved CNN-LSTM power load prediction model is proposed, first using the CNN feature extraction module to learn the local features of the input data; then input into the LSTM sequence learning model to extract the sequence feature information of the input data; at the same time, the self-attention mechanism is introduced into the LSTM to learn the features of the LSTM hidden layer, and the key special extraction is realized by assigning different attention weights to improve the final prediction accuracy; finally, ultra-short-term power load prediction, short-term power load prediction and medium- and long-term power load prediction are carried out, and the power consumption analysis before and after the epidemic, power consumption distribution analysis and power user portrait are assisted. The present invention can promote the digital upgrade of the power grid, meet the personalized needs of users, realize industry correlation analysis, dispatch power generation, predict power consumption trends, and guide resumption of work and production.

The specific implementation flow chart of the multi-factor power load forecasting method based on deep learning of the present invention as shown in Figure 1 includes: acquisition and storage of power load data and external environment data, data cleaning and preprocessing based on KNN and improved DBSCAN algorithm, construction and training based on improved CNN-LSTM model and power load forecasting. The specific implementation steps are as follows:

Step 1: Obtain load data of different regions, years, and electricity consumption categories and external environmental factors such as temperature, humidity, wind speed, etc., and store them in the sqlite database. The specific steps are as follows:

1) Load data of different regions, years, and electricity consumption categories are obtained from channels such as the State Grid user electricity consumption information release platform. The data categories include user number (desensitized), region, urban network/rural network, user classification (such as high voltage, charging pile, grid-connected high voltage self-provided power plant, low voltage non-residential, low voltage resident, etc.), electricity consumption category (household electricity, commercial electricity, urban electricity, rural electricity, etc.), customer category (normal electricity customers, changed electricity customers), voltage level (ultra-high voltage, high voltage, low voltage), industry category, contract capacity, operating capacity, first power supply date, frozen electricity date and 15-minute frozen electricity, etc. Table 1 shows the description of the electricity load data fields. The electricity load data is stored in the sqlite database to facilitate data migration and system deployment.

Table 1 Description of power load data fields

2) The acquisition and storage of data on external environmental factors such as temperature, humidity, and wind speed are completed by writing a Python crawler program. The Python crawler process involves simulated login, page acquisition and parsing, data structure design, and storage of external environmental data. Data on external environmental factors such as temperature, humidity, and wind speed are obtained from the official interface of the National Meteorological Website. The data categories include the current date, region, maximum temperature, minimum temperature, average humidity, average wind speed, weather type (the threshold range is set to [1,4], indicating the degree of cloudy to sunny), the current temperature of 15 minutes, and whether it is a holiday (the value is 0 or 1). Table 2 shows the description of the external environmental data fields. The time of the environmental data is synchronized with the time of the power load data. Similarly, the environmental data is stored in the sqlite database to facilitate data migration and system deployment.

Table 2 External environment data field description

Field Name Field meaning WEATHER_DATE Current Date WEATHER_ADDR Region MAX_TEMP Maximum temperature LOW_TEMP Minimum temperature AVERAGE_HUMIDITY Average humidity AVERAGE_WIND Average wind speed WEATHER_TYPE Weather type (threshold [1,4], indicating the degree of cloudy to sunny) R1_WEATHER Current temperature over 15 minutes IS_HOLIDAY Is it a holiday (value is 0 or 1)

Step 2: Cleaning and preprocessing of power load data and environmental impact data, including: abnormal data detection based on k-nearest neighbor (KNN) algorithm and improved DBSCAN algorithm, autoregressive interpolation and sequence data normalization. Refer to Figure 2 for a schematic diagram of the data processing flow of the present invention.

The specific steps are as follows:

1) Detection of abnormal power load data based on k-nearest neighbor (KNN) algorithm and improved DBSCAN algorithm. Due to power equipment failure or human factors, the acquired user power load data may contain a small amount of missing or abnormal data. These missing or abnormal data have a great impact on the prediction accuracy of user power load. Therefore, it is necessary to detect these missing or abnormal data and use methods such as elimination or interpolation to restore normal data values.

The present invention proposes an outlier detection method based on an improved KNN algorithm. The basic rule of the KNN algorithm is to find k neighboring samples or the most similar samples (1＜k＜N, N represents the total number of samples) of sample c in the feature space. If most of these k neighboring samples belong to a category, then c also belongs to this category. The average distance between sample c and its neighboring samples is calculated as the abnormality score of sample c. The higher the abnormality score, the more abnormal the sample c is. In order to better reflect the abnormal situation of the power load data, the present invention uses the total distance from each cluster to its respective k neighbors as the final abnormality score after obtaining the abnormality score of the power load data at each moment using the improved KNN algorithm. Then the set N _k (c _i ) of k neighbors of the power load data c _i at a certain moment i is expressed as:

表示c_i的k个邻近点之一，d_k(c_i)为c_i的k个邻近的平均距离，

表示c_i与其邻近点

的距离；则电负荷数据c_i的异常得分表示为：In the formula,

represents one of the k neighbors of _ci , d _k ( _ci ) is the average distance of the k neighbors of _ci ,

represents _ci and its neighboring points

The distance; then the abnormal score of the electric load data c _i is expressed as:

Where N _k ( _ci ) represents the set of k neighboring points of _ci ;

Finally, the top m clusters in the anomaly score sorted list are output as outliers of the power load data.

Since the KNN algorithm interferes with abnormal clusters and normal clusters when calculating the k neighbors of cluster c _i , it is easy to have misdetection and missed detection after the abnormal detection result. In order to solve this problem, the present invention proposes an improved clustering-based anomaly detection algorithm DBSCAN. The DBSCAN algorithm is a typical density-based clustering algorithm, which divides closely connected samples into different categories and finally obtains clustering category results, in which points that do not belong to any cluster are considered to be abnormal points. Compared with the traditional DBSCAN algorithm that uses globally unified parameters Eps and MinPst to achieve clustering, the present invention proposes an improved data partitioning method based on a sliding window, which first uses local parameters to achieve density clustering of small sample data, and then iteratively clusters the local clustering results to achieve the final global clustering results. The steps of the improved DBSCAN algorithm include three steps: parameter update, clustering and anomaly detection.

In the process of parameter update, we first need to set the size M of the cluster sliding window and calculate the average distance difference of the power load data in the window, set the first k neighboring power loads as MinPst, and set the Euclidean distance between the power load data as Eps. In order to alleviate the inconsistency between the power load data, a weight is set for each power load data to reduce the impact on the final clustering result. The calculation formula of the weight w( _ci , _cj ) is as follows:

Among them, Cov( _ci , _cj ) is the covariance of the power load data _ci at time i and the power load data _cj at time j, Var( _ci ) is the variance of the power load data _ci at time i, and Var( _cj ) is the variance of the power load data _cj at time j; r( _ci , _cj ) is expressed as the correlation coefficient of _ci and _cj , and the smaller the value, the greater the correlation; the power load data anomaly detection result is obtained by analyzing the clustering results. In the clustering process, the first marked abnormal point is set as a candidate abnormal point, and the abnormal score is set plus 1 (the initial value is 0). In the cyclic iterative clustering process, the next candidate abnormal point is entered and the abnormal score is updated. If the abnormal score S is equal to the number of clusters C, it is marked as an abnormal point.

2) For the processing of missing values of power load, interpolation is usually used to fill in the missing values. The present invention uses Lagrange interpolation for the missing power load data values. The idea of interpolation is to establish a suitable interpolation function f(x) based on known points. Therefore, the unknown point x _i is obtained by the interpolation function f(x) as f(xi), so that (x _i ,f(x _i )) can be used to approximate the missing point.

The idea of Lagrange interpolation ^{is to make the n-1 polynomial y=a0+a1x+a2x2} _{+ L +} a _n- ^1xn-1 pass through the coordinates of n points ( _x1 , _y1 ), ( _x2 , _y2 ), ( _x3 , _y3 ), L, ( _xn , _yn ), then the function expression of Lagrange interpolation can be expressed as:

Among them, _xi represents the i-th moment of the power user, _yi represents the power load of the power user at the i-th moment; n represents the total number of moments of power load.

3) Normalize the user power consumption data and environmental data. Since the data of different categories or the same category vary greatly, such as the power load data and temperature have great differences in numerical representation, and the power load data of different types of users also have great differences in numerical values, if normalization is not performed, it will have a great impact on the training of the model, and the influence of a certain parameter will be hidden or amplified. Therefore, the normalization formula for each indicator data is:

Among them, X={ _X1 , _X2 , _X3L , _XN } represents data of the same category, such as power load data, temperature data or wind speed, so the normalized value of the data is Y={ _Y1 , _Y2 , _Y3L , _YN }, and N is the total number of power load data or environmental impact data of a certain category.

Step 3: Construct a power load data prediction network model based on the improved CNN-LSTM. FIG3 is a schematic diagram of the power load prediction model based on the improved CNN-LSTM proposed by the present invention. CNN and LSTM are the most widely used deep learning technologies. The CNN model can be used to extract valuable features and filter out the noise of the input data, and the LSTM network can capture sequence pattern information. Although the LSTM network can process time-related sequence information, it only uses the attributes provided in the training set. In contrast, CNN can be used to extract local features and the same features that appear in different regions, but it does not have the characteristics of processing time series information. Therefore, a hybrid model that utilizes the advantages of the two deep learning technologies can improve the prediction accuracy. In addition, the present invention introduces a self-attention mechanism to improve CNN-LSTM, which can reconstruct the features of the hidden layer of LSTM.

Therefore, the specific steps are as follows:

Step 3.1) Preparation of training data for the power load prediction model, using the power load data and environmental data obtained in step 2) as training data, for example, taking short-term (daily) power load prediction as an example, the training data is the pre-processed data of a user's daily power load and environmental factors such as maximum temperature, minimum temperature, average temperature, average humidity, etc. for 365 days a year, and the power load prediction model is a neural network. First, load the data set and divide it into training set, verification set and test set according to 7:2:1, and then construct Dataloader for data reading for the training set and test set respectively. When training the model, only the data of each batch is calculated. The batch size loads the data into the memory and shuffles the order of each batch of data to improve the robustness of the model training. The batch size is set to 10 in the present invention.

Step 3.2) Construction of the power load forecasting model. The deep learning model in the present invention uses Python language experiment, and the improved CNN-LSTM power load forecasting model is implemented based on the Pytorch deep learning library. The improved CNN-LSTM power load forecasting model includes three parts: CNN feature learning module, LSTM sequence learning module and self-attention mechanism module.

The CNN feature learning module consists of three one-dimensional convolutional layers, in which the MaxPooling layer and the ReLu layer are added between two consecutive convolutional layers. The convolution operation is introduced to learn the features of the standardized power load data and environmental data. The feature map output by the convolutional layer has a limitation, that is, it will track the exact location of the input data features and can obtain more discriminative features in the input data. A MaxPooling layer is usually added after the convolutional layer to alleviate the limitation of the invariance of the generated feature map, and the activation function ReLu is used to enhance the model's ability to learn complex structures, thereby reducing the overall computational load and making the network easier to train.

When predicting the power load, the time series correlation of the load data should be fully considered. Compared with the traditional recurrent neural network, LSTM can accurately learn the long-term dependency in the time series and is suitable for learning long-period power load data. Therefore, the present invention uses LSTM to predict the power load and can obtain higher prediction accuracy. The input is the influencing parameter xt at time _t and the predicted value _ht-1 at the previous time. The predicted value _ht at time t is obtained through the prediction function F. Its function expression is as follows:

h _t =F(x _t ,h _t-1 )

LSTM enhances the ability of recurrent neural networks to handle long sequence dependency problems by adding memory and control gates, and shows better performance in predicting power load.

In the LSTM sequence learning module, the present invention uses three LSTM layers, each layer contains twenty neurons. The first two LSTM layers output the complete sequence of hidden states, and in the last LSTM layer, the last time step of the hidden state is output.

In order to fully explore the features of the hidden layer inside the LSTM for the final prediction ability of the model, the present invention adds a self-attention module between the three LSTMs in the sequence learning module. The self-attention mechanism assigns weights to the features of the hidden layer extracted by the LSTM, thereby mining more discriminative features of the power load data and environmental data to improve the final model prediction accuracy. The output sequence of the hidden layer of the penultimate layer of the LSTM network at time t is assigned the self-attention weight w _tl , which is expressed as:

Among them, L _h is the length of the sequence output by the LSTM hidden layer, l represents the sequence number of the LSTM hidden layer output sequence, and s _tl represents the direct similarity between the sequence l of the LSTM hidden layer at time t and other sequences;

The feature h _tl of sequence l is multiplied by its corresponding weight w _tl to form a new feature sequence h _t ' _l and input it into the next LSTM. Introducing the self-attention mechanism to assign weights to the features of the hidden layer of LSTM can effectively obtain the key information in the sequence, thereby improving the accuracy and efficiency of prediction.

In the development of any deep learning model, the dropout layer involves randomly selecting neurons and deactivating some of them during the training process to prevent overfitting of the model. In the present invention, a dropout layer is added between the CNN feature extraction block and the LSTM sequence learning to prevent overfitting. The output of the LSTM sequence learning block is also connected to a dropout layer, followed by a fully connected layer (FC) to produce the final output.

Step 3.3) Define the optimizer and set the model training parameters. The ratio of training data, validation data, and test data in the model training input data is 7:2:1, and the mean absolute error (MAE) is used as the loss function to monitor the validation loss. The optimizer of the model is set to the adaptive optimizer Adam, and the initial learning rate is set to 0.001. Adam can adaptively update the learning rate during the training process, making the model converge faster and the adjustment of parameters easier. In addition, the number of training iterations of this model is set to 700, the initial loss function of each iteration is set to 0, and the batch size is set to 10.

Step 4: The prediction of power load mainly includes: ultra-short-term power load prediction, short-term power load prediction and medium- and long-term power load prediction. Refer to Figures 4, 5 and 6, which respectively show the comparison line graphs of ultra-short-term, short-term and medium- and long-term power load predictions with the actual power load of users.

Figure 4 (a) shows the comparison between the predicted value and the actual value of the commercial electricity load on May 1st in a certain place in 2019. By analyzing the actual value and predicted value of electricity consumption on a certain day in the place, it can be obtained that the predicted value of commercial electricity consumption on that day can basically reflect the actual electricity consumption situation, which can provide scheduling guidance for power plants. (b) is the comparison of the commercial electricity load on May 1st in the place in 2018/2019/2020. By analyzing the comparison of the commercial electricity load on May 1st in the place in 2018/2019/2020, it is found that the commercial electricity load on May 1st in 2020 was more affected by the epidemic than in the previous two years.

Figure 5 is a line graph comparing the short-term electricity load forecast value and the actual value of commercial electricity consumption in a certain area. The temperature factor is introduced to predict the commercial electricity consumption in this area in 2019. The forecast value can basically reflect the actual load electricity consumption trend and the change trend of load temperature. However, an abnormal situation occurred around February 5th because most of the businesses in this area were on holiday during the Spring Festival.

Figure 6 is a line chart comparing the medium- and long-term commercial electricity load forecast and the actual value. (a) is the weekly electricity load forecast for the area; by comparing the weekly commercial electricity load value and the forecast value, the forecast value can basically reflect the changing trend of the actual value. (b) is the monthly electricity load forecast for the area; by comparing the actual value and the forecast value of the monthly commercial electricity load, the forecast value can basically reflect the changing trend of the actual value.

The specific steps are as follows:

Taking the commercial electricity consumption of a certain city as an example, the design and implementation of the power load forecasting module mainly predicts the power load from the ultra-short-term, short-term and medium-term dimensions from the level of the power plant power generation department, so as to facilitate the power plant power generation department to grasp the commercial electricity consumption of a certain city every day, every week and every month, and to make effective power load dispatch. For example, the power load data of the commercial electricity consumption time, day, week and month of a certain city in 2019, and environmental data such as the highest temperature, lowest temperature, average humidity, average wind force, weather type, etc. of each day are obtained. After data preprocessing and standardization, they are used as the input of the LSTM model, and the predicted power load trends of the commercial electricity consumption time, day, week and month of a certain city in 2019 are output. By comparing the real value and the predicted value, the accuracy of the model in predicting the power load is reflected, and the comparison of the changing trend of factors such as temperature and the power consumption trend is introduced to reflect the impact of environmental factors on the power load size.

The present invention is mainly used in the power sector, such as guiding power plants for power generation and dispatching, assisting power management departments in analyzing power consumption, mining user power consumption information, and can be applied to research and development departments to mine user power consumption patterns and improve the work efficiency of power departments. Reference Figure 7 shows the application scheme of the present invention. Factors that affect the power load of the power grid include region, time, temperature, power type, voltage type, etc. This work designs a multi-factor power load forecasting system based on deep learning. First, the data is preprocessed and standardized using abnormal data detection, autoregressive interpolation and sequence data normalization based on the k-nearest neighbor (KNN) algorithm and the improved DBSCAN algorithm; then an improved CNN-LSTM power load forecasting model is proposed. First, the CNN feature extraction module is used to learn the local features of the input data; then it is input into the LSTM sequence learning model to extract the sequence feature information of the input data; at the same time, the self-attention mechanism is introduced into the LSTM to learn the features of the LSTM hidden layer, and the key special extraction is achieved by assigning different attention weights to improve the final prediction accuracy; finally, ultra-short-term power load forecasting, short-term power load forecasting and medium- and long-term power load forecasting are carried out, and power consumption analysis before and after the epidemic, power consumption distribution analysis and power user portraits are assisted. The present invention can be applied to the power department to predict the power consumption trend of users, reasonably plan the phased power consumption, help the power department to dispatch the power generation of power plants, and can help guide the resumption of work and production in the industry and the correlation analysis of the industry.

Claims (5)

1. A multi-factor power load prediction method based on deep learning is characterized by comprising the following steps:

step 1: acquiring power load data including different areas, years and power utilization categories and environmental influence data including temperature, humidity and wind power, and storing the data in a sqlite database;

step 2: the method for cleaning and preprocessing the power load data and the environmental impact data comprises the following steps: abnormal data detection, autoregressive interpolation and sequence data normalization based on a k-proximity algorithm and an improved DBSCAN algorithm;

and step 3: introducing a self-attention mechanism, performing feature reconstruction on the hidden layer of the LSTM to realize end-to-end deployment, constructing an electrical load data prediction neural network model based on the improved CNN-LSTM, and using the electrical load data and the environmental impact data obtained by processing in the step (2) as a training set and a test set;

the step 3 specifically includes:

step 3.1: for the loading and preparation of network model data:

loading a data set, dividing a training set, a verification set and a test set, constructing a Dataloader as a data reader for the training set and the test set respectively, calculating data of each batch during model training, loading the data into a memory by the Dataloader according to the size of the batch, and disordering the data of each batch to improve the robustness of the model training;

step 3.2: constructing an electrical load data prediction neural network model based on the improved CNN-LSTM:

the neural network model comprises a CNN feature learning module, an LSTM sequence learning module and an attention mechanism module:

1) The CNN feature learning module comprises three one-dimensional convolution layers, wherein a Max painting layer and a ReLu layer are added between two continuous convolution layers; learning the characteristics of the normalized power load data and the normalized environmental impact data through convolution operation, and using the characteristics as a characteristic diagram output by the convolution layer; adding a MaxPolling layer to relieve the limitation of invariance of the generated feature diagram, and activating a function ReLu to enhance the capability of the model to learn a complex structure;

2) The LSTM sequence learning module comprises three LSTM layers, each layer comprising twenty neurons; the first two LSTM layers output the complete sequence of the hidden state, and the last LSTM layer outputs the last time step of the hidden state;

the input of the LSTM layer is an influencing parameter x at the time t _t And the predicted value h of the previous moment _t-1 Obtaining the predicted value h at the time t through the prediction function F _t (ii) a The function expression is as follows:

h _t ＝F(x _t ,h _t-1 )

3) A self-attention mechanism module is respectively added among the three LSTM layers and distributes weight to the features of the hidden layer extracted from the LSTM layer, so that electricity is used for miningLoad data and environmental impact data have more discriminative characteristics; hidden layer output sequence of penultimate layer of LSTM network at t moment is distributed from attention weight w _tl The expression is as follows:

in the formula, L _h For the sequence length of the LSTM hidden layer output, l denotes the sequence number of the LSTM hidden layer output sequence, s _tl Representing the direct similarity of the sequence l of the LSTM hidden layer at the time t and other sequences;

the feature h of the sequence l _tl Weight w corresponding thereto _tl Multiplying to form a new signature sequence h _t ' _l And input into the next LSTM;

step 3.3: defining an optimizer, setting model training parameters:

monitoring verification loss by using the average absolute error as a loss function, and setting a self-adaptive optimizer Adam to adaptively update the learning rate in the training process; setting training iteration times epoch, an initial loss function and the size of batch;

and 4, step 4: predicting the electric load through an improved electric load data prediction neural network model of the CNN-LSTM;

the electric load data comprises a user number, a region to which the user belongs, an urban/rural network, a user classification, an electricity utilization type, a client type, a voltage grade, an industry type, a contract capacity, an operation capacity, a first power transmission date, a frozen electric quantity date and a frozen electric quantity of 15 minutes; the environmental impact data comprises the current date, the region, the highest temperature, the lowest temperature, the average humidity, the average wind power, the weather type, the current temperature of 15 minutes and whether the current temperature is a holiday or not; and the time of the environmental impact data is synchronized with the time of the electrical load data.

2. The method for predicting the electrical load with the multiple factors based on the deep learning according to claim 1, wherein in the step 2, the abnormal data detection based on the k-neighbor algorithm and the improved DBSCAN algorithm is specifically as follows:

step 2.1: defining the average distance between a sample and an adjacent sample as the abnormal score of the sample, obtaining the abnormal score of the electricity load data at each moment by using a modified KNN algorithm, and taking the total distance from each cluster to each k adjacent samples as the final abnormal score;

load data c of electricity consumption at a certain time i _i K adjacent sets N of _k (c _i ) Expressed as:

in the formula (I), the compound is shown in the specification,

denotes c _i One of the k neighboring points of (a), d _k (c _i ) Is c _i K adjacent average distances of->

Denotes c _i And its adjacent point->

The distance of (d);

electrical load data c _i Is expressed as:

in the formula, N _k (c _i ) Is shown by c _i K sets of neighboring points;

finally, outputting the first m clusters of the abnormal score ranking list as abnormal values of the electric load data;

step 2.2: adopting an improved cluster-based anomaly detection algorithm DBSCAN, firstly utilizing local parameters to realize density clustering of data of small samples, then carrying out iterative clustering on local clustering results to realize a final global clustering result, and marking points which do not belong to any cluster class and belong to anomaly points; the method specifically comprises the following steps:

step a) updating parameters:

setting the size M of a cluster sliding window, calculating the average distance difference of the electric load data in the window, setting the former k adjacent electric loads as MinPst, and setting the Euclidean distance between the electric load data as Eps;

setting a weight for each load data to reduce the impact on the final clustering result, weight w (c) _i ,c _j ) The calculation formula of (c) is as follows:

wherein, cov (c) _i ,c _j ) Electrical load data c at time i _i And the electric load data c at the time of j _j Covariance of (a), var (c) _i ) Electrical load data c at time i _i Variance of (c), var (c) _j ) Electrical load data c for time j _j The variance of (a); r (c) _i ,c _j ) Is denoted by c _i And c _j The smaller the value, the greater the correlation;

step b) obtaining an abnormal detection result of the electric load data by analyzing the clustering result:

in the clustering process, setting the abnormal point marked for the first time as a candidate abnormal point, setting the abnormal score plus 1, entering the next candidate abnormal point in the cyclic iterative clustering process, and updating the abnormal score; if the anomaly score S is equal to the cluster number C, the anomaly point is marked.

3. The method for predicting the multi-factor electrical load based on the deep learning of claim 1, wherein in the step 2, the autoregressive interpolation is specifically as follows:

completing the missing electric load data value by adopting a Lagrange interpolation method so that the polynomial y = a of n-1 ₀ +a ₁ x+a ₂ x ² +…+a _n-1 x ^n-1 Coordinate (x) passing through n points ₁ ,y ₁ ),(x ₂ ,y ₂ ),(x ₃ ,y ₃ ),...,(x _n ,y _n ) Then the functional expression for lagrange interpolation is expressed as:

in the formula, x _i And x _j I and j time instants, y, representing the power consumers, respectively _i A power load indicating an i-th time of the power consumer; n represents the total number of times of the electrical load.

4. The method for predicting the electrical load according to claim 1, wherein in the step 2, the processing formula of the normalization of the sequence data is as follows:

wherein X = { X ₁ ,X ₂ ,X ₃ ...,X _N Denotes the same category of electrical load data or environmental impact data; the normalized value is Y = { Y = { (Y) } ₁ ,Y ₂ ,Y ₃ …,Y _N }，l∈[1,N]And N is the total number of certain types of electrical load data or environmental impact data.

5. The method for predicting the multifactor electric load based on the deep learning of claim 1, wherein in the step 4, the prediction of the electric load comprises: ultra-short term power load prediction, and medium-long term power load prediction.

Images