CN111178585A - Fault reporting amount prediction method based on multi-algorithm model fusion - Google Patents

Abstract

The embodiment of the application provides a fault report prediction method based on multi-algorithm model fusion, which comprises the steps of obtaining fault report data, extracting report data based on characteristics, and inputting the report data into a combined model; training each model in the combined model, and adjusting the weight value corresponding to the output result of each model based on the training result; and obtaining a receiving quantity predicted value output by the combined model based on the adjusted weight value. The method adopts multi-model fusion algorithm design including a cyclic neural network LSTM based on a deep learning technology, an XGboost algorithm and a LightGBM algorithm based on a gradient lifting tree and a traditional time series algorithm SARIMA algorithm, and adopts a weighting mode to obtain a final prediction result, so that the accuracy of prediction can be improved.

Description

Fault reporting amount prediction method based on multi-algorithm model fusion

Technical Field

The invention belongs to the field of data prediction, and particularly relates to a fault reporting quantity prediction method based on multi-algorithm model fusion.

Background

As the first-aid repair time is more easily influenced by various artificial and objective factors, compared with the fault report receiving amount, the first-aid repair time is more difficult to predict. Factors influencing the fault rush-repair duration are numerous, including environmental factors (weather conditions such as temperature), rush-repair personnel factors (rush-repair team, rush-repair personnel and the like), occurrence time, the position of the fault, the fault type and the like. In order to predict the repair duration more accurately, it is necessary to analyze factors affecting the repair duration.

Disclosure of Invention

In order to solve the defects and shortcomings in the prior art, the invention provides a fault report prediction method based on multi-algorithm model fusion, and a report prediction is carried out by constructing a combined model formed by a plurality of models, so that a more accurate prediction result can be obtained, the preparation can be carried out in advance according to the prediction result, and the error probability is reduced.

Specifically, the method for predicting fault reporting volume based on multi-algorithm model fusion provided by the embodiment of the present application includes:

acquiring fault report receiving data, extracting the report receiving data based on characteristics, and inputting the report receiving data into a combined model;

training each model in the combined model, and adjusting the weight value corresponding to the output result of each model based on the training result;

and obtaining a receiving quantity predicted value output by the combined model based on the adjusted weight value.

Optionally, the obtaining fault report receiving data, performing feature-based extraction on the report receiving data, and inputting the report receiving data into the combined model includes:

determining input features;

and performing feature extraction on the fault reporting data based on the determined input features.

Optionally, the determining the input features includes:

processing the initial features to obtain combined features;

long-term features are calculated across the day,

and combining the known characteristics to obtain the final input characteristics.

Optionally, the training each model in the combined model, and adjusting the weight value corresponding to the output result of each model based on the training result, includes:

a seasonal ARIMA model SARIMA is used as a time series prediction model;

obtaining main parameters of the LightGBM through a parameter searching method;

obtaining main parameters of the XGboost through a parameter searching method;

obtaining the main parameters of the LSTM by a parameter searching method;

after the three algorithms are trained respectively, the output results of the three algorithms need to be fused according to the weights.

Optionally, the method for predicting the fault reporting volume further includes:

constructing an average absolute error MAE as a performance evaluation index of the combined model,

wherein y is_iActual value, f, representing the output of the ith combined model_iAnd (4) representing a predicted value corresponding to the ith actual value, wherein the value ranges of n and i are positive integers.

The technical scheme provided by the invention has the beneficial effects that:

the method adopts multi-model fusion algorithm design including a cyclic neural network LSTM based on a deep learning technology, an XGboost algorithm and a LightGBM algorithm based on a gradient lifting tree and a traditional time series algorithm SARIMA algorithm, and adopts a weighting mode to obtain a final prediction result, so that the accuracy of prediction can be improved.

Drawings

In order to more clearly illustrate the technical solution of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.

Fig. 1 is a schematic flow chart of a fault reporting quantity prediction method based on multi-algorithm model fusion according to an embodiment of the present application.

Detailed Description

To make the structure and advantages of the present invention clearer, the structure of the present invention will be further described with reference to the accompanying drawings.

Example one

The fault reporting volume prediction method based on multi-algorithm model fusion, as shown in fig. 1, includes:

11. acquiring fault report receiving data, extracting the report receiving data based on characteristics, and inputting the report receiving data into a combined model;

12. training each model in the combined model, and adjusting the weight value corresponding to the output result of each model based on the training result;

13. and obtaining a receiving quantity predicted value output by the combined model based on the adjusted weight value.

In the implementation, the fault report receiving amount is a time series which changes with time by taking a day as a unit, so the prediction of the fault report receiving amount also belongs to the prediction problem of the time series. The prediction of the fault first-aid repair time is a typical regression prediction problem. The prediction of the time series can be roughly divided into two methods, one is a traditional time series analysis method, and the other is a method for converting the time series into a supervised machine learning problem; for the latter, a machine learning model can be established for prediction from the perspective of supervised learning. In fact, factors which often influence the change of the time series data are uncertain or incomplete, and all the influencing factors cannot be found out in reality, so that the established model is not accurate enough. More generally, only the predicted variables themselves are available, in which case they can no longer be translated into supervised machine learning problems, which require processing using conventional time series analysis models. The project adopts a traditional time series analysis method and a method combining supervised machine learning, and various models are used for modeling.

Based on the service understanding and data exploration of fault report volume, the method of character coding, one-hot vector, normalization and the like is adopted for data cleaning, the consistency of the data is checked to improve the data quality, and the main weather data, holidays and other factors are analyzed, so that 26 kinds of characteristic indexes including initial characteristics, combined characteristics, differential characteristics and the like are extracted in a combination mode, a differential mode and the like. Because the prediction task is complex, in order to obtain a better prediction effect, the project adopts multi-model fusion algorithm design. The models used in the method comprise a recurrent neural network LSTM based on a deep learning technology, an XGboost algorithm and a LightGBM algorithm based on a gradient lifting tree and a traditional time series algorithm SARIMA algorithm, and the final prediction result is obtained in a weighting mode.

Optionally, the obtaining fault report receiving data, performing feature-based extraction on the report receiving data, and inputting the report receiving data into the combined model includes:

determining input features;

and performing feature extraction on the fault reporting data based on the determined input features.

The weather factors influencing the report volume analyzed in the previous part comprise: daytime temperature, nighttime temperature, daytime precipitation, nighttime precipitation, daytime wind power, nighttime wind power, daytime weather conditions, nighttime weather conditions and the like, and in addition, factors such as whether the conditions are holidays or not are included, and a specific characteristic list is shown in table 1.

Initial characteristics	Combined features	Differential characterization
			Fault reporting volume	Mean temperature	Fault reporting quantity-difference
Daytime temperature	Temperature difference	Daytime temperature-differential
			Temperature at night	All day precipitation	Temperature difference at night
Precipitation in daytime	Mean wind force	Daytime precipitation-difference
			Precipitation at night	-	Precipitation-difference at night
Daytime wind power	-	Daytime wind-differential
			Wind power at night	-	Night wind-differential
Daytime weather conditions	-	Fault reporting quantity-difference
			Weather conditions at night	-	Mean temperature-difference
Holiday	-	Temperature difference-difference
			-	-	Total daily precipitation-difference
-	-	Mean wind-differential

Table 1 data major sources

The first step of feature extraction requires processing of the initial features to obtain combined features. For example, the average temperature and the temperature difference can be obtained from the daytime temperature and the nighttime temperature, and the total day precipitation can be obtained from the daytime precipitation and the nighttime precipitation. And the second step of calculating the long-term features across days, such as the difference features of the first two days of features, and finally combining the three features to obtain the final input feature.

The forecast report volume is in days, the historical data of multiple days can be used for forecasting the data of one day in the future, and the historical data of multiple days can be input besides the available features of the current day. The project models the data of three days before using, the data of three days are respectively processed for the final characteristics, the data of three days are spliced together to be used as the final input characteristics,

optionally, the training each model in the combined model, and adjusting the weight value corresponding to the output result of each model based on the training result, includes:

a seasonal ARIMA model SARIMA is used as a time series prediction model;

obtaining main parameters of the LightGBM through a parameter searching method;

obtaining main parameters of the XGboost through a parameter searching method;

obtaining the main parameters of the LSTM by a parameter searching method;

after the three algorithms are trained respectively, the output results of the three algorithms need to be fused according to the weights.

Because the model relates to the fusion problem of the three algorithms, the training model needs to optimize the three algorithms respectively. The tuning parameters of the three algorithms are given below.

SARIMA：

Due to seasonal influence factors existing in the prediction of fault report quantity, a seasonal ARIMA model SARIMA is used as a time series prediction model and is expressed as SARIMA (P, D, Q) (P, D, Q) s. Here, (P, D, Q) are the above non-seasonal parameters, while (P, D, Q) follow the same definition, but apply to the seasonal component of the time series. s is the period of the time series (quarterly 4, annually 12, etc.). The grid searching technology is used for performing traversal optimization work on the hyper-parameters. Finally, (P, D, Q) ═ 1,1, (P, D, Q, s) ═ 0,1,1,4, are obtained.

LightGBM：

The main parameter table of the LightGBM obtained by the parameter search method is as follows:

model parameters	Remarks for note	Value taking
			boosting_type	Booster type	gbdt
objective	Objective function	regression
			metric	Evaluation index	mae
num_leaves	Maximum number of leaves	32
			learning_rate	Learning rate	0.01
feature_fraction	Randomly selecting feature ratios	0.9
			bagging_freq	bagging frequency	3
verbose	Log output	-2

Table 2 LightGBM parameter setting table

XGBoost：

The main parameters of the obtained XGBoost by the parameter search method are shown in table 2:

table 2 XGBoost parameter setting table

LSTM：

The main parameters of the LSTM obtained by the parameter search method are shown in table 3:

network layer	Super parameter setting	Output size
			Input	-	-
Bidirection LSTM	units＝64	(,64)
			Dropout	0.5	(,64)
Dense	1	(,1)

TABLE 3 data Primary sources

After the three algorithms are trained respectively, the output results of the three algorithms need to be fused according to the weights. Here, based on the results of multiple experiments, the LightGBM, SARIMA, XGBoost and LSTM are analyzed to have the weights: 0.65, 0.1, 0.15.

Optionally, the method for predicting the fault reporting volume further includes:

constructing an average absolute error MAE as a performance evaluation index of the combined model,

wherein y is_iActual value, f, representing the output of the ith combined model_iAnd (4) representing a predicted value corresponding to the ith actual value, wherein the value ranges of n and i are positive integers.

After the algorithm model is established, evaluation is needed to judge the quality of the model. The model is typically built using a training set and evaluated using a test set. For the regression prediction task, the commonly used performance evaluation indexes include Mean absolute Error MAE (Mean absolute Error), Mean Square Error RMSE (Root Mean Square Error), MAPE (Mean absolute percentage Error), R-Square, and the like. The project data is complex, various abnormal data exist, so that RMSE and MSE which are more sensitive to the abnormal data are not selected to be used, and the average absolute error MAE is used as a main performance evaluation index.

Table 4 shows the performance comparison of the three algorithms. As can be seen from the table, the XGboost effect in a single algorithm is the best; the weighted results after the multi-model fusion are carried out are good in average result. Because the prediction of the duration takes 15 minutes as a unit, the MAE of the prediction result is 1.15, and the corresponding error of the emergency repair duration is about 17.25 minutes.

Type of algorithm	Mean absolute error MAE
		LightGBM	1.52
Xgboost	1.21
		LSTM	1.43
Average result	1.28
		Weighted result	1.15

TABLE 4 comparison of algorithmic Performance

The sequence numbers in the above embodiments are merely for description, and do not represent the sequence of the assembly or the use of the components.

The above description is only exemplary of the present invention and should not be taken as limiting the invention, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (5)

1. The fault reporting volume prediction method based on multi-algorithm model fusion is characterized by comprising the following steps:

acquiring fault report receiving data, extracting the report receiving data based on characteristics, and inputting the report receiving data into a combined model;

training each model in the combined model, and adjusting the weight value corresponding to the output result of each model based on the training result;

and obtaining a receiving quantity predicted value output by the combined model based on the adjusted weight value.

2. The method for predicting fault report based on multi-algorithm model fusion according to claim 1, wherein the obtaining fault report data, the feature-based extraction of the report data and the inputting of the report data into the combined model comprises:

determining input features;

and performing feature extraction on the fault reporting data based on the determined input features.

3. The fault exposure prediction method based on multi-algorithm model fusion as claimed in claim 2, wherein the determining input features comprises:

processing the initial features to obtain combined features;

long-term features are calculated across the day,

and combining the known characteristics to obtain the final input characteristics.

4. The method of claim 1, wherein the training of each model in the combined model and the adjusting of the weight value corresponding to the output result of each model based on the training result comprises:

a seasonal ARIMA model SARIMA is used as a time series prediction model;

obtaining main parameters of the LightGBM through a parameter searching method;

obtaining main parameters of the XGboost through a parameter searching method;

obtaining the main parameters of the LSTM by a parameter searching method;

after the three algorithms are trained respectively, the output results of the three algorithms need to be fused according to the weights.

5. The fault report prediction method based on multi-algorithm model fusion according to claim 1, further comprising:

constructing an average absolute error MAE as a performance evaluation index of the combined model,

wherein y is_iActual value, f, representing the output of the ith combined model_iAnd (4) representing a predicted value corresponding to the ith actual value, wherein the value ranges of n and i are positive integers.

Images