Power load prediction method and system based on improved DBN-LSTM

Abstract

The invention belongs to the technical field of power load prediction, and particularly discloses a power load prediction method and system based on improved DBN-LSTM, and the method comprises the following steps: constructing an improved DBN prediction model and an improved LSTM prediction model; based on the multiple linear regression model, distributing weights of the DBN prediction model and the LSTM prediction model, and combining prediction results to obtain a DBN-LSTM prediction model; training the DBN-LSTM prediction model, and adjusting the weight of the DBN prediction model and the weight of the LSTM prediction model; and utilizing the trained DBN-LSTM prediction model to obtain a final power load prediction value. According to the power load prediction method and system based on the improved DBN-LSTM provided by the invention, the advantages of the DBN prediction model and the LSTM prediction model are complementary by establishing the combined model of the improved DBN-LSTM, so that the prediction effect is relatively accurate, and relatively good robustness is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a power load forecasting method and system based on improved DBN-LSTM. Background Art

[0002] The management of electricity costs is particularly important, because electricity is not only one of the key factors in steel production, but also accounts for a large proportion of the total production cost. Therefore, monitoring and analyzing the electricity consumption of the steel industry can help companies more clearly understand their electricity consumption under different energy use methods and the corresponding costs, thereby encouraging them to optimize their energy use behavior, helping them to better plan and manage energy use, and providing reasonable energy-saving and emission reduction suggestions, thereby reducing overall costs.

[0003] The invention patent with application publication number CN 118868064 A discloses a combined power load forecasting method, which can greatly save the workload of data processing by training a multi-scale feature extraction CNN LSTM network based on high-frequency nonlinear components. However, since the model is a forecasting model composed of different substructures, the timeliness of the forecast is reduced, the storage parameters occupy more memory, and the training time is long. Therefore, it is necessary to simplify the forecasting model. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a power load forecasting method and system based on improved DBN-LSTM. By establishing an improved DBN-LSTM combined model, the advantages of the two prediction models, DBN and LSTM, are complemented, so that the prediction effect is more accurate and has better robustness.

[0005] The present invention adopts the following technical solution.

[0006] In the first aspect, the present invention provides an electric load forecasting method based on an improved DBN-LSTM, comprising the following steps: constructing an improved DBN forecasting model and an LSTM forecasting model respectively; based on a multivariate linear regression model, allocating weights of the DBN forecasting model and the LSTM forecasting model and combining the forecasting results to obtain a DBN-LSTM forecasting model; training the DBN-LSTM forecasting model and adjusting the weights of the DBN forecasting model and the LSTM forecasting model; and using the trained DBN-LSTM forecasting model to obtain a final electric load forecasting value.

[0007] Preferably, the constructing of the improved DBN prediction model comprises the following steps: constructing the DBN prediction model; and optimizing the DBN prediction model using an improved Adam algorithm.

[0008] Preferably, the iterative process using the improved Adam algorithm is as follows:

[0009] m t =γm t-1 +g t

[0010] ν t =βν t-1 +(1-β)m t

[0011] Among them, γ is the learning rate; t is the number of iterations; L(θ t ) is the loss function in the target optimization problem; is L(θ t ) Gradient with respect to θ; first-order variable m t Equivalent to E(g t ), that is, the current gradient g t The expectation of; 0≤β<1 is a hyperparameter; the second-order variable v t Equivalent to E(g t 2 ), which is the square of the current gradient g t 2 expectations.

[0012] Preferably, the DBN-LSTM prediction model is:

[0013] L=λ 0 +λ DBN L DBN +λ LSTM L LSTM +μ

[0014] Among them, λ 0 is a constant; DBN and λ LSTM are the weights of the DBN prediction model and the LSTM prediction model; L DBN and L LSTM are the predicted load values ​​of the DBN prediction model and the LSTM prediction model respectively; μ is the random error.

[0015] Preferably, after constructing the improved DBN prediction model and LSTM prediction model, the following steps are also included: dividing the data set in the collected historical prediction samples into a training data set and a test data set, and dividing the test data set into a first test data set and a second test data set; training the constructed DBN prediction model and LSTM prediction model with the training data set, and testing the trained DBN prediction model and LSTM prediction model on the first test data set and the second test data set respectively to obtain the DBN prediction load value and the LSTM prediction load value.

[0016] Preferably, the training of the DBN-LSTM prediction model and adjusting the weights of the DBN prediction model and the LSTM prediction model comprises the following steps: using the obtained DBN prediction load value and the LSTM prediction load value and the first test data set to train the constructed DBN-LSTM prediction model;

[0017] The trained DBN-LSTM prediction model is tested on the second test data set, the weights of the DBN and LSTM models are adjusted, and the prediction accuracy of the DBN-LSTM prediction model is ensured to meet the judgment conditions.

[0018] Preferably, the judgment condition is as follows: using the prediction evaluation index to calculate the error value between the predicted value and the actual value obtained by the DBN prediction model, the LSTM prediction model and the DBN-LSTM prediction model respectively; if the error value of the DBN-LSTM prediction model is smaller than the error value of the DBN prediction model and the error value of the LSTM prediction model, the judgment condition is considered to be met.

[0019] Preferably, the prediction evaluation indicators are mean absolute percentage error and root mean square error.

[0020] Preferably, the formula for the mean absolute percentage error is:

[0021]

[0022] The formula for the root mean square error is:

[0023]

[0024] in, and Y t are the predicted value and actual value of the power load obtained by the prediction model at time t, and n is the number of prediction points.

[0025] In a second aspect, the present invention provides a prediction system using the aforementioned power load prediction method based on improved DBN-LSTM, the prediction system comprising: a data acquisition module, a DBN prediction module, an LSTM prediction module and a DBN-LSTM prediction module; the data acquisition module is used to collect historical prediction data of time-of-use electricity prices; the DBN prediction module is used to construct a DBN prediction model based on historical prediction data, and obtain the power prediction value predicted by the DBN prediction model; the LSTM prediction module is used to construct an LSTM prediction model based on historical prediction data, and obtain the power prediction value predicted by the LSTM prediction model; the DBN-LSTM prediction module is used to assign weights of the DBN prediction model and the LSTM prediction model based on a multivariate linear regression model, construct a DBN-LSTM prediction model, and obtain the final power prediction value using the DBN-LSTM prediction model.

[0026] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides an improved DBN-LSTM power load forecasting method, which is aimed at the problem that the daily power consumption data of the steel industry under the time-of-use electricity price has a strong time series correlation, and among the commonly used forecasting methods, DBN has a strong feature learning ability but is slightly insufficient in processing long time series and medium-term load forecasting, and LSTM has high prediction accuracy but also has high requirements for data and resources. In order to solve the problem of low robustness of a single prediction model, the present invention complements the advantages of the two prediction networks, establishes a combined model based on the improved DBN-LSTM, constructs a DBN prediction model and an LSTM prediction model respectively, and then establishes a multivariate linear regression model to find the linear relationship between the load prediction value and the true value of a single model, assigns appropriate weights to a single DBN prediction model and a LSTM prediction model, and linearly combines the prediction results to obtain an improved DBN-LSTM combined model, and finally uses the combined model to forecast power load, thereby exploring the changing trend of power consumption of steel enterprises under different electricity price scenarios and evaluating the effect of the time-of-use electricity price policy under different market conditions. The results show that DBN-LSTM is more accurate than single LSTM and DBN networks in predicting power load for the same data set for two days, and the prediction deviation does not increase with the increase of prediction time, which shows good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a DBN-LSTM model prediction flow chart in the power load prediction method based on improved DBN-LSTM in the present invention;

[0028] Figure 2 This is a flow chart of DBN-LSTM model training in the power load forecasting method based on improved DBN-LSTM in the present invention;

[0029] Figure 3 It is a comparison curve chart of the predicted value obtained by using the DBN-LSTM prediction model in the present invention and the actual value and the predicted value obtained by using the LSTM prediction model alone. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0031] like Figure 1 As shown, embodiment 1 of the present invention provides a power load forecasting method based on an improved DBN-LSTM, by respectively establishing DBN (Deep Belief Nets) and LSTM (Long Short-Term Memory) prediction models, combining DBN and LSTM networks by using a non-uniform weighting method based on multivariate linear regression, constructing an improved DBN-LSTM model, and improving the robustness of the model; simulating and analyzing the development trend of power load in the steel industry in Jiangsu Province under time-of-use electricity prices, specifically comprising the following steps:

[0032] Step 1: Build an improved DBN prediction model and an LSTM prediction model respectively, and predict the DBN prediction load value and the LSTM prediction load value through the DBN prediction model and the LSTM prediction model. Step 1 specifically includes the following steps:

[0033] Step 1.1: Build a DBN prediction model.

[0034] A deep belief network model is constructed for prediction. Deep Belief Network (DBN) is a deep learning model with a composite architecture consisting of multiple hidden layers. It can be trained in two stages: unsupervised layer-by-layer pre-training and supervised reverse fine-tuning.

[0035] Step 1.2: Use the improved Adam algorithm to optimize the DBN prediction model;

[0036] Aiming at the problem that DBN prediction results are greatly affected by parameters, the improved Adam algorithm is used to optimize the network.

[0037] Adam adds a primary variable m before iterating the independent variable. tTo calculate the exponential weighted average of past gradients, Adam is adjusted and simplified. The iterative process is as follows:

[0038] m t =γm t-1 +g t

[0039] ν t =βν t-1 +(1-β)m t

[0040] Among them, γ is the learning rate; t is the number of iterations; L(θ t ) is the loss function in the target optimization problem; is L(θ t ) Gradient with respect to θ; first-order variable m t Equivalent to E(g t ), that is, the current gradient g t The expectation of; 0≤β<1 is a hyperparameter; the second-order variable v t Equivalent to E(g t 2 ), which is the square of the current gradient g t 2 expectations.

[0041] Using parameter m t 、v t Adjust the learning of each element in the gradient matrix to calculate the learning rate corresponding to each gradient parameter. The iterative process is:

[0042]

[0043] Among them, η is the learning rate; ε is the stability constant, the goal of the stability constant is to avoid the denominator being 0.

[0044] Step 1.3: Build an LSTM prediction model.

[0045] A long short-term memory network model is constructed for prediction. The long short-term memory network (LSTM) is a time recurrent neural network. Each neuron contains three control gates for cyclic update: input gate, forget gate and output gate, which can better avoid the problem of gradient disappearance.

[0046] Step 1.4, dividing the data set in the collected historical prediction samples into a training data set and a test data set, and dividing the test data set into a first test data set test1 and a second test data set test2;

[0047] Step 1.5: Use the training data set to train the constructed DBN prediction model and LSTM prediction model, and test the trained DBN prediction model and LSTM prediction model on the first test data set test1 and the second test data set test2 respectively to obtain the DBN prediction load value and the LSTM prediction load value.

[0048] Step 2: Based on the unequal weight method of multivariate linear regression, the advantages of DBN and LSTM networks are combined to build a DBN-LSTM prediction model.

[0049] DBN can more effectively identify and handle outliers in power load forecasting, but this ability may also affect the overall forecasting accuracy to a certain extent. In addition, as the forecasting time increases, the forecasting deviation of DBN usually increases gradually, resulting in a decrease in the reliability of long-term forecasting. Although the forecasting model based on the deep LSTM network has high forecasting accuracy, the forecasting stability is not high, and large forecasting errors will occur at certain time points. Therefore, in order to further improve the forecasting accuracy of the forecasting model, combining the advantages of DBN and LSTM networks and making up for the low robustness of a single model, a power load forecasting method based on the DBN-LSTM combined model is proposed.

[0050] However, due to the different prediction accuracy of DBN and LSTM, and the different proportions of importance of each independent variable in the dependent variable, the average weight combination prediction model cannot be used to combine the prediction results of the two methods. Therefore, a weight allocation method based on multiple linear regression is adopted. The prediction results of the DBN and LSTM models and the actual load values ​​are input into the linear regression model as input data. The linear regression model is used to analyze the relationship between the prediction results and the actual load values, and appropriate weights are assigned to each prediction result. Finally, the prediction results are combined based on the assigned weights to obtain the final load prediction value. The non-average weight model based on multiple linear regression combined prediction can be expressed as:

[0051]

[0052] In the formula, i is the independent variable, L is the dependent variable, and λ i Represents the independent variable L i The weight coefficients are , and the weight coefficients satisfy the sum of 1.

[0053] The multivariate linear regression model not only assigns weights to each method, but also generates a constant term similar to a neuron bias and a random error term to adjust the final combined prediction value. Therefore, the DBN-LSTM prediction model based on the non-equal weight method of multivariate linear regression is constructed as follows:

[0054] L=λ 0 +λDBN L DBN +λ LSTM L LSTM +μ

[0055] In the formula, λ 0 is a constant; DBN and λ LSTM are the weights of DBN and LSTM models respectively; L DBN and L LSTM are the predicted load values ​​of DBN and LSTM models respectively; μ is the random error.

[0056] Step 3: Train the DBN-LSTM prediction model and adjust the weights of the DBN and LSTM models in the DBN-LSTM prediction model, such as Figure 2 Step 3 specifically includes the following steps:

[0057] Step 3.1, using the DBN predicted load value and LSTM predicted load value obtained in step 1 and the first test data set test1, train the constructed DBN-LSTM prediction model;

[0058] Step 3.2: Test the DBN-LSTM prediction model trained in step 3 on the second test data set test2, adjust the weights of the DBN prediction model and the LSTM prediction model, and ensure that the prediction accuracy of the DBN-LSTM prediction model meets the judgment conditions.

[0059] Furthermore, the judgment conditions are as follows:

[0060] The prediction evaluation index is used to calculate the error between the predicted value and the actual value obtained by the DBN prediction model, the LSTM prediction model and the DBN-LSTM prediction model;

[0061] If the error value of the DBN-LSTM prediction model is smaller than the error value of the DBN prediction model and the error value of the LSTM prediction model, it is considered that the judgment condition is met.

[0062] Among them, the prediction evaluation indicators include Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE).

[0063] Step 4: Use the trained DBN-LSTM prediction model to obtain the combined power load prediction value.

[0064] Step 5: Analyze the development trend of power load in the steel industry under the adjustment of time-of-use electricity prices through simulation, such as Figure 3 shown.

[0065] Taking Jiangsu Province as an example, when predicting the power load of the steel industry in Jiangsu Province under the adjustment of time-of-use electricity prices, it is necessary not only to analyze the internal change law of the load, but also to consider the impact of electricity price factors on the load. Since the power load has continuity, similarity and working characteristics, taking the load at time t on the forecast day a as an example, the selected input of the traditional load forecasting model is shown in Table 1.

[0066] Table 1 Prediction model input

[0067]

[0068]

[0069] The time-of-use electricity price data of Jiangsu Province from July to August 2023 is used as the data set, and the data set is divided into a training set train and a test set test. Because the model is composed of multiple models, the test set needs to be further divided, and different models need to be trained and tested separately. First, the DBN and LSTM models are trained with the training set train, and then the trained DBN and LSTM models are tested with the test set test1 to obtain the prediction results of the two models. Then, the two prediction results and the actual load in the test set test1 are input into the multiple linear regression model to train the model, and the training of the entire model is completed. Finally, the model is tested with the test set test2 to obtain the final combined prediction value. Its multiple linear regression calculation parameters are shown in Table 2.

[0070] Table 2 Weight coefficients of the multivariate linear regression model

[0071]

[0072] In order to evaluate the prediction performance of the model, MAPE and RMSE are used as evaluation indicators. MAPE is a relative error measure that uses absolute values ​​to avoid the offset of positive and negative errors. Relative errors can be used to compare the accuracy of predictions of various time series models. RMSE is the square root of the mean square error (MSE) and is used to provide the standard deviation of the prediction error. In theory, the smaller the values ​​of MAPE and RMSE, the better the prediction model fits and has better accuracy. The formulas for MAPE and RMSE are as follows:

[0073]

[0074] In the formula, and Y t are the predicted value and actual value of load at time t respectively, and n is the number of prediction points.

[0075] Table 3 shows the prediction errors of the DBN-LSTM combined model and the two single models of DBN and LSTM. As can be seen from the table, in terms of prediction accuracy, the prediction accuracy of the combined prediction model is the highest, the accuracy of the LSTM prediction model is second, and the prediction accuracy of the DBN network is the lowest. This low accuracy is probably due to the dense representation and gradient diffusion of the DBN network, which is exactly why it is necessary to combine other algorithms such as the LSTM network to improve the overall stability. Although the prediction accuracy of the LSTM network is higher than that of DBN, LSTM will produce prediction deviations as the prediction time increases, so the accuracy is not ideal. For the DBN-LSTM combined prediction, this improved method can complement the advantages of the two and make up for their shortcomings. Therefore, it is better in terms of prediction accuracy and stability. The prediction error of the DBN-LSTM combined model is lower than that of both the DBN and LSTM models, and the MAPE values ​​are improved, which shows the robustness of the DBN-LSTM combined model.

[0076] Table 3 Comparison of prediction errors of DBN-LSTM combined model and DBN and LSTM single models

[0077]

[0078] The beneficial effect of the present invention is that, compared with the prior art, the present invention aims at the problem that the daily electricity consumption data of the steel industry under the time-of-use electricity price has a strong time series correlation, and among the commonly used prediction methods, DBN has a strong feature learning ability but is slightly insufficient in processing long time series and medium-term load prediction, and LSTM has high prediction accuracy but also has high requirements for data and resources. In order to solve the problem of low robustness of a single prediction model, the present invention complements the advantages of the two prediction networks, establishes a combined model based on an improved DBN-LSTM, constructs a DBN prediction model and an LSTM prediction model respectively, and then establishes a multivariate linear regression model to find the linear relationship between the load prediction value and the true value of a single model, assigns appropriate weights to a single DBN prediction model and a LSTM prediction model, and linearly combines the prediction results to obtain an improved DBN-LSTM combined model, and finally uses the combined model to predict power load, thereby exploring the changing trend of electricity consumption of steel enterprises under different electricity price scenarios and evaluating the effect of the time-of-use electricity price policy under different market conditions. The results show that DBN-LSTM is more accurate than single LSTM and DBN networks in predicting power load for the same data set for two days, and the prediction deviation does not increase with the increase of prediction time, which shows good robustness.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A power load forecasting method based on improved DBN-LSTM, characterized in that: The following steps are involved: Construct improved DBN prediction model and LSTM prediction model respectively; Based on the multivariate linear regression model, the weights of the DBN prediction model and the LSTM prediction model are assigned and the prediction results are combined to obtain the DBN-LSTM prediction model; Train the DBN-LSTM prediction model and adjust the weights of the DBN prediction model and the LSTM prediction model; The trained DBN-LSTM prediction model is used to obtain the final power load prediction value.

2. The power load forecasting method based on improved DBN-LSTM according to claim 1 is characterized in that: The improved DBN prediction model is constructed, comprising the following steps: Build DBN prediction model; The improved Adam algorithm is used to optimize the DBN prediction model.

3. The power load forecasting method based on improved DBN-LSTM according to claim 2 is characterized in that: The iterative process of the improved Adam algorithm is as follows: m t =γm t-1 +g t n t =bn t-1 +(1-β)m t Among them, γ is the learning rate; g t =▽ θ L(θ t ), t is the number of iterations; L(θ t ) is the loss function in the target optimization problem; ▽ θ L(θ t ) is L(θ t ) Gradient with respect to θ; first-order variable m t Equivalent to E(g t ), that is, the current gradient g t The expectation of; 0≤β<1 is a hyperparameter; the second-order variable v t Equivalent to E(g t 2 ), which is the square of the current gradient g t 2 expectations.

4. The power load forecasting method based on improved DBN-LSTM according to claim 1 is characterized in that: The DBN-LSTM prediction model obtained is: L=λ0+λ DBN L DBN +λ LSTM L LSTM +μ Among them, λ0 is a constant; λ DBN and λ LSTM are the weights of the DBN prediction model and the LSTM prediction model; L DBN and L LSTM are the predicted load values ​​of the DBN prediction model and the LSTM prediction model respectively; μ is the random error.

5. The power load forecasting method based on improved DBN-LSTM according to claim 1 is characterized in that: After constructing the improved DBN prediction model and the LSTM prediction model, the following steps are also included: Dividing the data set in the collected historical prediction samples into a training data set and a test data set, and dividing the test data set into a first test data set and a second test data set; The constructed DBN prediction model and LSTM prediction model are trained with the training data set, and the trained DBN prediction model and LSTM prediction model are tested on the first test data set and the second test data set respectively to obtain the DBN prediction load value and the LSTM prediction load value.

6. The power load forecasting method based on improved DBN-LSTM according to claim 5 is characterized in that: The training of the DBN-LSTM prediction model and the adjustment of the weights of the DBN prediction model and the LSTM prediction model include the following steps: Using the obtained DBN predicted load value and LSTM predicted load value and the first test data set, the constructed DBN-LSTM prediction model is trained; The trained DBN-LSTM prediction model is tested on the second test data set, the weights of the DBN and LSTM models are adjusted, and the prediction accuracy of the DBN-LSTM prediction model is ensured to meet the judgment conditions.

7. The power load forecasting method based on improved DBN-LSTM according to claim 6 is characterized in that: The judgment conditions are as follows: The prediction evaluation index is used to calculate the error between the predicted value and the actual value obtained by the DBN prediction model, the LSTM prediction model and the DBN-LSTM prediction model; If the error value of the DBN-LSTM prediction model is smaller than the error value of the DBN prediction model and the error value of the LSTM prediction model, it is considered that the judgment condition is met.

8. The power load forecasting method based on improved DBN-LSTM according to claim 7 is characterized in that: The prediction evaluation indicators are mean absolute percentage error and root mean square error.

9. The power load forecasting method based on improved DBN-LSTM according to claim 8 is characterized in that: The formula for the mean absolute percentage error is: The formula for the root mean square error is: in, and Y t are the predicted value and actual value of the power load obtained by the prediction model at time t, and n is the number of prediction points.

10. A prediction system using the power load prediction method based on the improved DBN-LSTM according to any one of claims 1 to 9, characterized in that: The prediction system includes: a data acquisition module, a DBN prediction module, an LSTM prediction module and a DBN-LSTM prediction module; The data acquisition module is used to collect historical forecast data of time-of-use electricity prices; The DBN prediction module is used to build a DBN prediction model based on historical prediction data and obtain a power prediction value predicted by the DBN prediction model; The LSTM prediction module is used to build an LSTM prediction model based on historical prediction data and obtain a power prediction value predicted by the LSTM prediction model; The DBN-LSTM prediction module is used to allocate weights of the DBN prediction model and the LSTM prediction model based on a multivariate linear regression model, construct a DBN-LSTM prediction model, and obtain a final power prediction value using the DBN-LSTM prediction model.

Images