CN111598225B - Air conditioner cold load prediction method based on self-adaptive deep confidence network - Google Patents

Abstract

The invention discloses an air conditioner cold load prediction method based on a self-adaptive deep confidence network, which comprises the steps of firstly collecting cold load data, adopting a Lagrange interpolation method to compensate for missing and abnormal energy consumption data, and carrying out normalization processing on the processed energy consumption data; the processed data is processed through independent Gaussian distribution, and the processed CRBM is input to a prediction model; training is carried out through an RBM non-supervision mechanism, an RBM hidden layer of a previous layer is used as a visual layer input of an RBM of a next layer, and network parameters are regulated; then reverse training is carried out through a supervised BP neural network, and network parameters are regulated again; then adopting Adam optimization algorithm to adjust network parameters; finally, selecting parameters and structures of the prediction model, and carrying out structure evaluation selection on the prediction model by adopting a reconstruction error RE; and evaluating the result by adopting a root mean square relative error (RMSPE) and an average absolute percentage error (MAPE) to finish the air conditioner cold load prediction. The method has good prediction precision, universality and applicability.

Description

Air conditioner cold load prediction method based on self-adaptive deep confidence network

Technical Field

The invention belongs to the technical field of building air conditioner cold load prediction, and particularly relates to an air conditioner cold load prediction method based on a self-adaptive deep confidence network.

Background

With the increasing urbanization, industry development and population growth, global energy demand continues to rise. It is counted that building operating cold loads account for approximately 40% of global cold loads, with greenhouse gas emissions accounting for 1/3 of the total emissions, indicating that buildings have become the largest energy consumer. For buildings, the heating ventilation and air conditioning system occupies the largest proportion in the total cold load of the building, and the huge energy consumption increases the pressure of the power grid. Therefore, the energy management of the building is particularly important.

The prediction of the building cold load is an important part in the building energy management process, is a key work for realizing building energy conservation, correctly and reasonably predicts the building cold load, can timely and accurately discover some abnormal conditions or potential equipment faults in the building cold load, and is convenient for management personnel to take measures in time so as to avoid excessive waste of energy. Meanwhile, the accurate and reasonable prediction of the building cold load can provide a certain basis for the management personnel to reasonably distribute energy, so that the energy is reasonably and effectively used. And the power generation scheme can be reasonably arranged, so that the supply and demand balance of the power grid is realized, and the power system can stably operate, which is also a very important ring for the power system.

The artificial neural network model, the support vector machine model, the decision tree model and the mixed model in the existing cold load prediction method are widely applied. However, these models suffer from various problems that make the prediction accuracy less than ideal. For example, the learning speed of the artificial neural network algorithm is not high enough, and local optimization and over-fitting phenomena are easy to generate; the support vector machine is difficult to use in a large number of samples, and the problem of multi-classification is difficult to solve; the scheme probability of the decision tree model is easily affected by human factors, and the decision accuracy is reduced; and the problems of high calculation amount and high calculation difficulty of the mixed model and the like. The above prediction algorithm only focuses on the prediction of the self algorithm on the cold load, but cannot fully consider the complex characteristics of the cold load of the building, so that the prediction effect is not ideal.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides the air conditioner cold load prediction method based on the self-adaptive deep confidence network, which considers complex factors influencing the cold load of the building, realizes the effective prediction of the cold load condition of the building, achieves the technical effects of energy conservation and consumption reduction by controlling the influencing conditions, and solves the problems in the prior art.

The invention adopts the following technical scheme:

the air conditioner cold load prediction method based on the self-adaptive deep confidence network comprises the steps of firstly collecting cold load data, adopting a Lagrange interpolation method to compensate the missing and abnormal energy consumption data, and carrying out normalization processing on the processed energy consumption data; the processed data is processed through independent Gaussian distribution, and the processed CRBM is input to a prediction model; training by using an RBM non-supervision mechanism, taking an RBM hidden layer of a previous layer as a visual layer input of an RBM of a next layer, and adjusting network parameters theta= { w, a, b }; then, reverse training is carried out through a supervised BP neural network, and network parameters theta= { w, a, b } are regulated again; then adopting an Adam optimization algorithm to adjust network parameters theta= { w, a, b }; finally, selecting parameters and structures of the prediction model, and carrying out structure evaluation selection on the prediction model by adopting a reconstruction error RE; and evaluating the result by adopting a root mean square relative error (RMSPE) and an average absolute percentage error (MAPE) to finish the air conditioner cold load prediction.

Specifically, the prediction model adds continuous values of independent Gaussian distribution in the linear unit to simulate real data, and the energy function E (v, h; θ) is as follows:

wherein θ= { w, a, b, σ }, σ _i For visual layer v _i Standard deviation of corresponding gaussian noise;

the activation probabilities of the update visual layer and the hidden layer are:

wherein N (μ, σ) ² ) Representing the mean μ and variance σ of the gaussian function ² ，a _i H is the offset of the visual layer _j To hide the binary state of cell j, w _ij For the weight between the two, n is the number of the predicted cold load of the visual layer, v _i Inputting the binary state of i for the visual layer, b _j Is the offset of the hidden layer.

Specifically, in the training of the RBM non-supervision mechanism, the RBM converts an input vector from a visual layer to a hidden layer through an activation function, from the hidden layer to the visual layer, and extracts features through training to minimize an internal energy function, so as to obtain a joint configuration energy function E (v, h; theta) between the visual layer and the hidden layer, ensure function distribution standardization through joint distribution of the visual layer and the hidden layer, calculate conditional probability distribution according to the determined visual layer and the hidden layer, and determine an optimal model through maximizing a likelihood function by using a ladder-lifting method, and when a set of training sample sets S= { v is given ¹ ,v ² ,...v ⁿ ) And when the objective function is calculated as the maximized lower log likelihood function Ls.

Further, the log-likelihood function Ls under maximization is:

wherein P (v) ⁿ ) For the activation probability of each input sample, N is the number of training samples.

Further, the joint distribution between the visual layer and the hidden layer is:

P(v,h|θ)＝exp(-E(v,h))/Z(θ)

wherein Z (θ) is a partitioning function, v is a visual layer unit, h is a hidden layer unit, and after the visual layer and the hidden layer are determined, the conditional probability distribution is calculated as follows:

wherein a is _i H is the offset of the visual layer _j To conceal the binary state of cell j, n is the predicted amount of cooling load for the visual layer, w _ij For the weight between the two, m is the number of hidden layers corresponding to the visual layer, v _i Inputting the binary state of i for the visual layer, b _j Is the offset of the hidden layer.

Specifically, the readjustment network parameter θ= { w, a, b } is specifically:

inputting input vectors containing time, temperature, humidity and solar radiation into a network model, performing unsupervised network training, processing into continuous values through Gaussian distribution CRBM, and performing model training;

and dividing the training samples into f groups of training samples, training the f groups of training samples to adjust network parameters theta= { w, a, b } of the prediction model, and storing the network parameters after reaching the training layer number of the prediction model.

Specifically, the Adam optimization algorithm is adopted to adjust the network parameters θ= { w, a, b } specifically:

initializing the parameter vector, the first moment vector and the second moment vector, then updating each part in a loop and iteration mode to enable the parameter theta to be converged, namely, adding 1 to the time t, correspondingly updating each parameter of the deviation, and finally updating the parameter theta of the prediction model by using the calculated parameter value.

Specifically, the root mean square relative error RMSPE and the mean absolute percentage error MAPE are used for evaluation, the difference between the predicted value and the actual value of the predicted model is compared, the CRBM-DBN network model is evaluated by adopting a reconstruction error, the deviation between the true value and the predicted value forms the reconstruction error, the depth of the predicted model is calculated by calculating the reconstruction error, the number of hidden nodes is calculated, the MAPE and the training time are used as evaluation standards, and the cold load data is predicted by training.

Further, the root mean square relative error RMSPE and the mean absolute percentage error MAPE are specifically:

wherein y is _i And

the actual load value and the load predicted value at the ith moment are respectively;

The average value of the air conditioner cold load true value is k, and k is the number of samples of all test sets; the reconstruction error was used for the evaluation as follows:

wherein RE is a reconstruction error, X represents a min-batch matrix, aV _i Representing the state value of the current predicted visual element.

Further, the depth of the prediction model is 2, the number of hidden layer nodes is 20, and the prediction step length is 1h.

Compared with the prior art, the invention has at least the following beneficial effects:

the invention provides an air conditioner cold load prediction method based on a self-adaptive deep confidence network, which aims at solving the problem that building cold load data often has nonlinearity and dynamic characteristics, so that the cold load data cannot be reliably and accurately predicted. The algorithm combines the advantages of supervised learning and unsupervised learning, digs high-dimensional features, extracts feature values from bottom to top, and effectively solves a series of problems caused by random initialization of parameters of the traditional neural network. In addition to having a good initiation point, the problems of over-fitting and under-fitting that often occur with neural networks are also effectively solved by pre-training.

Furthermore, the selected data is subjected to deletion and abnormal data compensation processing, so that the problem of data reliability can be effectively solved, and in order to avoid prediction errors caused by different orders of magnitude of parameters of an input layer, the input data is subjected to normalization processing, so that the prediction errors of a network model caused by the data errors can be effectively solved.

Furthermore, the Boltzmann machine is processed by Gaussian distribution, so that the problem that the RBM can only accept binary input and data is easy to lose can be effectively solved.

Furthermore, the deep confidence network optimized by adopting the Adam algorithm has strong self-adaptability and good performance in the aspect of nonlinear problem treatment due to consideration of the second derivative of the objective function, and the problem of slow convergence speed in the process of training network model parameters is effectively solved.

Further, by evaluating the network model by using the reconstruction errors, the size of the reconstruction errors generated by selecting different structures can be clearly seen, so that parameters and structures of the model can be better selected.

In conclusion, the method accurately predicts the building cold load, has prediction precision higher than that of a BP neural network prediction model, has good prediction precision, universality and applicability, is particularly suitable for large public buildings with periodically-changed cold load, and provides cold load prediction data which are more useful for energy-saving planning and energy-saving planning.

The technical scheme of the invention is further described in detail through the drawings and the embodiments.

Drawings

FIG. 1 is an exception data handling diagram;

FIG. 2 is a schematic view of an RBM;

FIG. 3 is a schematic diagram of a CRBM-DBN network prediction;

FIG. 4 is a flow chart of the CRBM-DBN algorithm;

FIG. 5 is a graph of model reconstruction errors, wherein (a) is the reconstruction error at a network depth of 1, (b) is the reconstruction error at a network depth of 2, (c) is the reconstruction error at a network depth of 3, and (d) is the reconstruction error at a network depth of 4;

fig. 6 is a graph of the prediction results, wherein (a) is a graph of the prediction of the cooling load on the rest day and the working day, (b) is an effect of enlarging for 7 to 13 hours, and (c) is an effect of enlarging for 22 to 28 hours.

Detailed Description

The invention discloses an air conditioner cooling load prediction method based on a self-adaptive deep confidence network, which comprises the following steps of:

s1, collecting cold load data (including temperature, humidity, solar load, wind speed, time and the like), adopting a Lagrange interpolation method to compensate the missing and abnormal energy consumption data, and carrying out normalization processing on the processed energy consumption data to be used as cold load energy consumption prediction;

s101, a model aims at a market cold load prediction method, and the market load is 08:00-22:00, recorded for 15 hours. Assuming that the ith data is missing or anomalous during the day, a Lagrangian function is constructed using the remaining 14 known load points during the day:

wherein j is _j For the load value at the j-th time point, pj (x) is a polynomial of degree 14,

B _k for the remaining 14 time pointsCoordinate set, B _k E {1, 2..i.1, i+1,..14 }, the load value at the i-th time of the day is y _i ＝L ₁₄ (i)。

Referring to fig. 1, the missing and abnormal data is compensated by using a lagrangian interpolation method, and the original abnormal data and the results before and after the missing data are processed.

S102, in view of the fact that the orders of magnitude of the parameters in the input layer are different, in order to avoid prediction errors generated by the orders, the input parameters are normalized, and a specific formula is as follows:

wherein x' is the normalization result, x is the input vector including temperature, humidity, solar load, time, x _min And x _max Is the minimum and maximum of the corresponding samples in the input data.

S2, the data processed in the step S1 are firstly processed through independent Gaussian distribution, and the processed CRBM is input into a prediction model to ensure the continuity of the data;

the prediction model adds continuous values of independent Gaussian distribution to the linear unit to simulate real data, so that a traditional limited Boltzmann machine can process continuous input vectors, and an energy function E (v, h; theta) is as follows:

wherein θ= { w, a, b, σ }, σ _i For visual layer v _i Standard deviation of corresponding gaussian noise, typically σ ² There are better results when=1.

Thus, the activation probabilities of the updated visual and hidden layers are:

wherein N (μ, σ) ² ) Representing the mean μ and variance σ of the gaussian function ² 。

S3, training is carried out through an RBM non-supervision mechanism, an RBM hidden layer of a previous layer is used as a visual layer input of an RBM of a next layer, and network parameters theta= { w, a, b } are adjusted for the first time;

referring to fig. 2, the rbm converts an input vector from a visual layer to a hidden layer by activating a function, from the hidden layer to the visual layer, and extracts features by training a minimized internal energy function, where a joint configuration energy function between the visual layer and the hidden layer is:

wherein v is _i And h _j Is the binary state of the visual layer input i and hidden element j, a _i And b _j Offset, w, of the visual layer and the hidden layer, respectively _ij And m is the number of hidden layers corresponding to the visual layers, and n is the predicted cold load number of the visual layers.

The joint distribution between the visual layer and the hidden layer is:

P(v,h|θ)＝exp(-E(v,h；θ))/Z(θ)

where Z (θ) is a partitioning function to ensure that the distribution is normalized, v is the visual layer element, and h is the hidden layer element to ensure that the distribution of the function is normalized.

After the visual layer and the hidden layer are determined, the conditional probability distribution is calculated as follows:

the purpose of RBM training is to determine the best model by maximizing likelihood function using a bench-top approach, when given a set of training sample sets s= { v ¹ ,v ² ,...v ⁿ ) When the objective function is the maximum log likelihood function:

wherein P (v) ⁿ ) For the activation probability of each input sample, N is the corresponding number of input training samples.

After the training sample set is given, adjusting the network parameters theta= { w, a, b } can adjust the activation probability of the corresponding adjusting hidden layer so as to determine the state of the hidden layer; after the hidden layer is given, the activation probability of the corresponding visual layer can be adjusted by adjusting the network parameters theta= { w, a and b } so as to reconstruct the input data, namely, the activation probability of the hidden layer and the visual layer is maximized by maximizing the log likelihood function, and therefore, the reconstruction of the data is realized.

S4, performing reverse training through a supervised BP neural network, and adjusting network parameters theta= { w, a, b };

referring to fig. 3 and 4, input vectors (including time, temperature, humidity, solar radiation, etc.) are input into a network model, firstly, the model is trained through an unsupervised network, processed into continuous values through gaussian distribution CRBM, and the model is trained according to steps S2 and S3 from bottom to top;

dividing the training samples into f groups of small-batch training samples, training the f groups of training sets to adjust network parameters theta = { w, a, b }, and after the number of training layers of the model is reached, storing the trained network parameters for the next step;

the BP algorithm reversely supervises and finely adjusts network parameters;

s5, after input data are subjected to RBM non-supervision feature training and BP supervised training learning through Gaussian distribution processing, the input data X are converted into another feature space F (Y, theta) through the model, network parameters theta= { w, a, b } are calculated through minimizing errors of F (Y, theta) and X, and a mean square error function is:

where k is the number of test set cooling loads.

Adopting an Adam optimization algorithm to adjust the network parameter theta= { w, a, b };

the Adam optimization algorithm firstly initializes the parameter vector, the first moment vector and the second moment vector, then iteratively updates each part in a circulating way to enable the parameter theta to be converged, namely, t is added with 1, and the corresponding parameter of the updating deviation is updated, and finally the calculated parameter value is used for updating the parameter theta of the model.

The Adam optimization algorithm is adopted to minimize the loss function, the Adam algorithm designs independent self-adaptive learning rate for different parameters by calculating first moment estimation and second moment estimation of gradient, and has high calculation efficiency and lower memory requirement, and the process of updating network parameters is expressed as follows:

m _t ＝β ₁ ×m _t -1+(1-β ₁ )×g _t

wherein g _t Is the gradient of the mean square error function L (theta) to theta, m _t Is a first moment estimate of the gradient, n _t Is a second moment estimate of the gradient,

and->

Respectively to m _t And n _t Deviation correction of moment estimated exponential decay rate beta ₁ Is 0.9, beta ₂ A step size eta of 0.99 and a small constant epsilon of 10 with stable numerical value of 0.001 ^-8 。

S6, selecting parameters and structures of the model, and performing structure evaluation selection on the model by adopting a reconstruction error RE; the results were evaluated using the root mean square relative error RMSPE and the mean absolute percent error MAPE.

The most commonly used evaluation indexes, namely root mean square relative error RMSPE and mean absolute percentage error MAPE, are used for evaluation, and the difference between the predicted value and the actual value of the prediction model is compared, and the calculation formula is as follows.

Wherein y is _i And

the actual load value and the load predicted value at the ith moment are respectively;

And k is the number of samples of all test sets, and the smaller the RMSPE and MAPE values are, the better the prediction effect is.

Simple evaluation of the CRBM-DBN (Adam) network model using reconstruction errors (reconstruction error):

wherein RE is a reconstruction error, X represents a min-batch matrix, aV _i The state value of the current prediction visual unit is represented, the deviation between the true value and the predicted value forms a reconstruction error, and the depth of the model can be calculated by calculating the reconstruction error.

Referring to fig. 5, the setting of the structural parameters in the deep learning model affects the training process, training time, and prediction result. In order to optimize training to reduce errors, experiments are performed by increasing the number of RBM layers layer by layer, and calculating the reconstruction errors to determine the depth of the network, wherein when the depth of the network is 2 layers, the overall reconstruction errors are reduced and gradually tend to be stable, and the average value of the reconstruction errors is smaller, so that the accuracy is higher. When the network depth is other values, the average value of the reconstruction errors is larger, and the reconstruction errors are always in a fluctuation state. Thus, the network depth of the RBM is determined to be 2 layers, i.e., 2 hidden unit layers and 1 visual unit layer in total.

And for the number of hidden nodes, MAPE and training time are taken as evaluation criteria, and are selected through experiments.

Selection of hidden node number in Table 1

As can be seen from table 1, the model prediction effect is optimal when the number of hidden layer nodes is 20.

Therefore, the depth of the prediction model set by the method is 2, the number of hidden layer nodes is 20, the prediction step length is 1h, and the cold load data is predicted through training.

The predicted daily cooling load curve is generated, after parameters and structures of a network model are obtained, the predicted daily cooling load curve can be obtained, as shown in fig. 6, fig. 6 (a) is a load prediction result of different algorithms, fig. 6 (b) (c) is a load prediction result amplified in the peak load time period of the rest day and the working day, and as the Adam optimization algorithm considers the second derivative of the objective function, the model has strong adaptability and good performance in processing nonlinear problems, and as can be seen from the figure, the method provided by the invention has better prediction effect compared with other two prediction algorithms, and the prediction accuracy in the uncertain time periods such as peak load underestimation is higher.

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Prediction model result comparison analysis

The case carries out predictive analysis on holiday and workday cold load data:

TABLE 2 workday prediction model result analysis

TABLE 3 analysis of holiday prediction model results

The most commonly used evaluation index root mean square relative error RMSPE and mean absolute percentage error MAPE are used for evaluation, and the difference between the predicted value and the actual value of the prediction model is compared.

Table 4 model evaluation analysis

As can be seen from Table 4, the prediction accuracy and time complexity are improved greatly compared with the conventional DBN and CRBM-DBN methods.

The invention takes the business building with the complex western security as a research object, and establishes the cold load prediction model based on the daily cold load of the business building.

In summary, the invention provides an adaptive deep confidence network-based air conditioner cold load prediction method, which aims at the problem of data loss caused by that RBM can only accept binary input, introduces a CRBM model with Gaussian distribution and the problem of slow convergence rate of the CRBM-DBN when training network model parameters, and provides an adaptive learning CRBM-DBN model method, wherein in each iteration, the convergence rate is improved by updating objective functions on corresponding parameters theta (w, a, b) through updating objective functions. Experimental results show that the method provided by the method is better than the traditional DBN and CRBM-DBN in terms of prediction accuracy and time complexity, and is an effective air conditioner cold load prediction method.

The above is only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited by this, and any modification made on the basis of the technical scheme according to the technical idea of the present invention falls within the protection scope of the claims of the present invention.

Claims (7)

1. The air conditioner cold load prediction method based on the self-adaptive deep confidence network is characterized in that cold load data is collected firstly, missing and abnormal energy consumption data are subjected to compensation processing by adopting a Lagrange interpolation method, and the processed energy consumption data are subjected to normalization processing; the processed data is processed through independent Gaussian distribution, and the processed CRBM is input to a prediction model; training by using an RBM non-supervision mechanism, taking an RBM hidden layer of a previous layer as a visual layer input of an RBM of a next layer, and adjusting network parameters theta= { w, a, b }; then, reverse training is carried out through a supervised BP neural network, and network parameters theta= { w, a, b } are regulated again; then adopting an Adam optimization algorithm to adjust network parameters theta= { w, a, b }; finally, selecting parameters and structures of the prediction model, and carrying out structure evaluation selection on the prediction model by adopting a reconstruction error RE; the result is evaluated by adopting a root mean square relative error (RMSPE) and an average absolute percentage error (MAPE), and the air conditioner cold load prediction is completed;

in RBM non-supervision mechanism training, RBM converts input vector from visual layer to hidden layer by activating function, from hidden layer to visual layer, and extracts feature by minimizing internal energy function by training to obtain joint configuration energy function E (v, h; θ) between visual layer and hidden layer, and ensures function distribution standardization by joint distribution of visual layer and hidden layer, and calculates conditional probability distribution according to determined visual layer and hidden layer, and determines optimal model by maximizing likelihood function by using ladder-up method, when a set of training sample set S= { v is given ¹ ,v ² ,...v ⁿ ) When the objective function is calculated as a maximized lower log likelihood function Ls, the maximized lower log likelihood function Ls is:

wherein P (v) ⁿ ) The activation probability of each input sample is given, and N is the number of training samples;

the joint distribution between the visual layer and the hidden layer is:

P(v,h|θ)＝exp(-E(v,h))/Z(θ)

wherein Z (θ) is a partitioning function, v is a visual layer unit, h is a hidden layer unit, and after the visual layer and the hidden layer are determined, the conditional probability distribution is calculated as follows:

wherein a is _i H is the offset of the visual layer _j To conceal the binary state of cell j, n is the predicted amount of cooling load for the visual layer, w _ij For the weight between the two, m is the number of hidden layers corresponding to the visual layer, v _i Inputting the binary state of i for the visual layer, b _j Is the offset of the hidden layer.

2. The adaptive deep belief network-based air conditioner cold load prediction method of claim 1, wherein the prediction model adds continuous values of independent gaussian distributions to linear units to simulate real data, and the energy function E (v, h; θ) is:

wherein θ= { w, a, b, σ }, σ _i For visual layer v _i Standard deviation of corresponding gaussian noise;

the activation probabilities of the update visual layer and the hidden layer are:

wherein N (μ, σ) ² ) Representing the mean μ and variance σ of the gaussian function ² ，a _i H is the offset of the visual layer _j To hide the binary state of cell j, w _ij For the weight between the two, n is the number of the predicted cold load of the visual layer, v _i Inputting the binary state of i for the visual layer, b _j Is the offset of the hidden layer.

3. The adaptive deep belief network-based air conditioner cooling load prediction method according to claim 1, wherein the readjusting network parameter θ= { w, a, b } is specifically:

inputting input vectors containing time, temperature, humidity and solar radiation into a network model, performing unsupervised network training, processing into continuous values through Gaussian distribution CRBM, and performing model training;

and dividing the training samples into f groups of training samples, training the f groups of training samples to adjust network parameters theta= { w, a, b } of the prediction model, and storing the network parameters after reaching the training layer number of the prediction model.

4. The adaptive deep belief network-based air conditioner cold load prediction method according to claim 1, wherein the adjusting network parameters θ= { w, a, b } by Adam optimization algorithm is specifically:

initializing the parameter vector, the first moment vector and the second moment vector, then updating each part in a loop and iteration mode to enable the parameter theta to be converged, namely, adding 1 to the time t, correspondingly updating each parameter of the deviation, and finally updating the parameter theta of the prediction model by using the calculated parameter value.

5. The adaptive deep belief network-based air conditioner cold load prediction method according to claim 1, wherein the root mean square relative error RMSPE and the mean absolute percentage error MAPE are used for evaluation, the difference between the predicted value and the actual value of the prediction model is compared, the CRBM-DBN network model is evaluated by adopting the reconstruction error, the deviation between the actual value and the predicted value constitutes the reconstruction error, the depth of the prediction model is calculated by calculating the reconstruction error, and the cold load data is predicted by training with the MAPE and the training time as evaluation criteria for the number of hidden nodes.

6. The adaptive deep belief network-based air conditioner cold load prediction method according to claim 5, wherein the root mean square relative error RMSPE and the mean absolute percentage error MAPE are specifically:

wherein y is _i And

the actual load value and the negative value at the i-th moment respectivelyA load prediction value;

The average value of the air conditioner cold load true value is k, and k is the number of samples of all test sets; the reconstruction error was used for the evaluation as follows: />

Wherein RE is a reconstruction error, X represents a min-batch matrix, aV _i Representing the state value of the current predicted visual element.

7. The adaptive deep belief network-based air conditioner cold load prediction method according to claim 5, wherein the depth of the prediction model is 2, the number of hidden layer nodes is 20, and the prediction step length is 1h.

Images