Short-term power load prediction method based on MBBO-LSTM model

Abstract

The invention discloses a short-term power load prediction method based on an MBBO-LSTM model, and belongs to the technical field of power intelligent scheduling and decision support. According to the method, a time sequence coupled by power load historical data and multi-feature meteorological parameters is used as input; modeling the multi-dimensional associated information features by relying on a deep learning model integrated by a self-adaptive improved LSTM network model and an MBBO algorithm; carrying out model training through a scheme of combining a piecewise learning rate attenuation strategy and an advanced stopping mechanism; compared with other common deep learning methods based on the LSTM network under different climate conditions, the method shows excellent prediction performance, good generalization ability and high environmental adaptability, and is suitable for a short-term power load prediction task in a complex environment with multi-source information fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power intelligent scheduling and decision support, and particularly relates to a short-term power load prediction method based on an MBBO-LSTM model. BACKGROUND

[0002] Traditional power system operation highly depends on and consumes a large amount of limited fossil energy, which increases the amount of greenhouse gas emissions; in order to achieve the energy saving and emission reduction target and promote the energy structure transformation, it has become an inevitable trend to build a new type of power system mainly composed of renewable energy; however, the large-scale access of renewable energy significantly increases the volatility and operation uncertainty of the power system, and puts forward higher requirements for the control ability of the system; under this background, improving the accuracy of power load prediction has become a key means to ensure the efficient operation of the system and realize intelligent management.

[0003] In practical applications, the change trend of power load is not only affected by historical power consumption data, but also closely related to meteorological factors such as temperature, humidity and wind speed; in the new type of power system with increasing proportion of renewable energy, these meteorological variables have a decisive role on the stability of power load output; however, traditional prediction methods mostly rely on single historical load data to construct input features, ignore the dynamic influence of external environmental factors and are difficult to fully depict the spatio-temporal dependence relationship of power load, resulting in low prediction accuracy and weak generalization ability of the prediction model; therefore, it is necessary to integrate load historical data and multi-source meteorological information to construct input time series, so as to fundamentally improve the modeling capability of mathematical model on real environmental spatio-temporal characteristics; by fully revealing the multi-dimensional driving mechanism in the evolution process of power load, the prediction performance of the new type of power system can be improved; however, the increase of feature dimension significantly increases the difficulty and computational burden of prediction model design, which brings challenges to the development of efficient and accurate load prediction method.

[0004] There are great limitations in traditional methods for carrying out power load prediction based on high-dimensional and multi-feature power load-weather information sequences with strong time sequence dependence. For example, statistical models represented by autoregression (AR) and moving average (MA) are difficult to effectively capture the complex load characteristic change relationship. Machine learning methods represented by support vector regression (SVR), random forest (RF), extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM) have limited modeling capability for multi-dimensional input features, resulting in that the prediction accuracy is difficult to meet the actual demand. Therefore, researchers seek to improve the performance of this time series prediction by using deep learning technology with strong nonlinear modeling and feature extraction capability, especially LSTM network derived from recurrent neural network (RNN).

[0005] Studies have shown that traditional LSTM network has stronger modeling capability and higher prediction accuracy in the task of short-term power load prediction with single feature input. However, with the deep coupling of multi-source weather information such as temperature, humidity and air pressure with power load historical data, the modeling capability, prediction accuracy and training stability of the traditional LSTM network need to be improved or enhanced. Therefore, it is necessary to construct a new network model and propose a training mechanism and performance enhancement scheme to realize high-precision robust prediction of short-term power load in complex and real scenarios. SUMMARY

[0006] To solve the problem of high-precision robust prediction of short-term power load in a complex environment of multi-source information fusion, the present application proposes a short-term power load prediction method (MBBO-LSTM method) based on multi-batch Bayesian optimization (MBBO) enhanced LSTM network. The method aims to solve the following technical problems: (1) The traditional prediction method has limited feature modeling capability in the complex environment of coupling power load historical data and multi-source weather information, and cannot fully explore the driving mechanism of environmental factors on power load change; (2) In the face of complex multi-dimensional heterogeneous input time series, the traditional LSTM model is prone to performance instability or slow convergence in the training process, resulting in significant increase in operation cost; (3) The traditional LSTM model highly depends on manual experience to adjust the structure and training parameters, and is difficult to achieve optimal configuration, which restricts the further improvement of prediction performance. The technical scheme adopted by the present application is: a short-term power load prediction method based on an MBBO-LSTM model; specifically comprising the following steps: S1, collect historical power load data, multi-source meteorological data and time information, fuse into a multivariate time series, and generate a three-dimensional tensor sample set through sliding window sampling and normalization; divide the sample set into a training set, a validation set and a test set; The present application sets the power load-multisource meteorological time series obtained by actual observation as , wherein represents the length of the time series, F represents the number of variables contained in each time step in the time series; at t time step, the time series s is sampled using a sliding window with a length of w and a step size of X , to obtain an input sample ; the corresponding output sample is , which represents the load value at time t+s; select m consecutive samples to , to form a three-dimensional tensor sample set , wherein m represents the number of samples, i.e., the number of consecutive samples obtained by sliding window sampling from the original time series; at the same time, in order to reduce the dimensional difference between different attribute features, the input and output samples are scaled to the interval [0,1] using the min-max normalization method; When constructing the three-dimensional tensor sample set, the sliding window length s = 24h and the step size w = 1h are set, where h represents hours; the three-dimensional tensor sample set is divided into a training set and a test set according to a ratio of 8:2, and 20% of the samples in the training set are extracted as a validation set.

[0007] S2, construct an improved LSTM network model suitable for the input of the three-dimensional tensor sample set; input the training set, and model the multivariate time series features through the improved LSTM network model; the improved LSTM network model sequentially includes a multivariate input layer, a single LSTM layer, a Dropout layer, a fully connected layer and a regression output layer; The improved LSTM network model structure construction method of the present application is: The multivariate input layer receives a training set, a single LSTM layer uses a gating mechanism to establish a memory cell update process, a Dropout layer uses a regularization method to screen data features to reduce the risk of overfitting, a fully connected layer integrates data features and maps them to the output space of the prediction task, and a regression layer outputs the feature data of the training set after inverse normalization processing, thereby realizing the prediction of the future time change trend of the power load. The application sets each input sample as a three-dimensional tensor sample set χ with a length of s ; in each time step, the improved LSTM network updates its memory cell through a gating mechanism c t , thereby effectively modeling the time dependence in the entire sequence. First, the forgetting gate f t adjusts the retention of the memory of the previous time step, and the formula is as follows: ; wherein, x t is the current input, h t-1 is the hidden state of the previous moment, W f is the weight of the forgetting gate, b f is the bias term of the forgetting gate; Subsequently, the input gate i t is used to adjust the influence of the current input on the memory cell, and the formula is as follows: u t , , ; wherein, W i is the weight of the input gate, W u is the weight of the candidate state, b i is the bias term of the input gate, b u is the bias term of the candidate state, sigma(.) is a Sigmoid activation function, and tanh(.) is a hyperbolic tangent activation function. Then, the memory cell state c t is updated, and the formula is as follows: wherein, c ​t-1 is the memory state of the previous time step, and represents the element-wise multiplication (Hadamard) operation; Finally, the output gate o t controls the generation of the hidden state h t , which is given by: where W o is the weight of the output gate, b o is the bias term of the output gate; After the single LSTM layer, the final hidden state h T is input to a Dropout layer, which randomly shuts down some neurons during training to reduce overfitting and feature redundancy while preserving the propagation of temporal information; the output after the Dropout layer is input to a fully connected layer and a regression layer, and finally the predicted value of the power load data is output.

[0008] S3, the training scheme combining segmented learning rate decay and early stopping mechanism is adopted to dynamically adaptively train the improved LSTM network model; In the model training process, the learning rate is dynamically adjusted according to the change of the training round number, the segmented learning rate decay (PLRDS) is used to ensure that a higher learning rate is maintained in the early training to accelerate convergence, and then the learning rate is gradually reduced to realize stable optimization, the segmented learning rate decay η t is as follows: ; wherein η t is the learning rate in the training process, η 0 is the initial learning rate, t e is the current training round number, T d is the preset learning rate decay period, γ is the preset learning rate decay factor, is the floor operation; At the same time, the early stopping mechanism (ESM) is used to dynamically monitor the validation loss; every f v validation frequency monitors the validation loss once, if no performance improvement is observed in continuous P times of evaluation, the training process is terminated in advance, thereby effectively preventing overfitting and saving computing resources; In the present application, ηThe initial learning rate is set to 0.01. f v The verification frequency is set to 10. P Set to 100 times.

[0009] S4. The LSTM network model structure and training hyperparameters are optimized and improved in stages using the MBBO algorithm to complete the short-term power load forecasting task based on the MBBO-LSTM model; the hyperparameters include the number of hidden units, initial learning rate, learning rate decay factor, learning rate decay period and Dropout rate. This invention proposes an MBBO-LSTM model, such as Figure 2 As shown; the MBBO hyperparameter optimization algorithm is used to adjust the hyperparameters of the number of hidden units, initial learning rate, learning rate decay factor, learning rate decay period, and Dropout rate; the first stage optimizes the LSTM network structure parameters, including the number of hidden units and the initial learning rate, to improve feature representation capability and model convergence speed; in this invention, the number of hidden units is set to [10, 300], and the initial learning rate is set to [1×10]. -3 The second stage adjusts the learning rate decay factor to achieve a more stable training process; in this invention, the learning rate decay factor is set to [0.001-0.002] and the learning rate decay period is set to [120, 160]; the third stage fine-tunes the Dropout rate to alleviate overfitting and improve the model's generalization ability; in this invention, the fine-tuned Dropout rate is [0.01, 0.025]; each stage reconstructs the surrogate objective function based on the optimal result of the previous stage and further refines the search in the updated hyperparameter space; This invention proposes an improved LSTM network structure and a set of hyperparameters for training as { θ 1, θ 2, ···, θ n},in n The number of hyperparameters. n The value is 5; based on the performance metrics on the validation set. E RMSE As the objective function f ( θ The hyperparameter optimization problem can be formalized as follows: ; in, θ * The optimal combination of hyperparameters that minimizes the objective function is argmin, which represents the input that minimizes the objective function. The MBBO algorithm proposed in this invention uses a Gaussian process to evaluate the objective function. f ( θ) construct a probabilistic surrogate model and estimate the posterior distribution of the optimal hyperparameter combination through observed samples to predict the function value under any value input: wherein, μ ( θ ) is the predicted mean, k ( θ , θ ′) is the covariance function, which is used to describe the correlation and uncertainty between samples; The present application adopts expected improvement (EI) as the core acquisition function of the MBBO strategy, and balances exploration and utilization when selecting new sampling points: , ; wherein, f min (θ) is the currently observed optimal target value, μ ( θ ) is the predicted mean of the surrogate model, σ ( θ ) is the standard deviation of the surrogate model, Φ( z ) is the cumulative distribution function of the standard normal distribution, φ ( z ) is the probability density function of the standard normal distribution, E (·) is the expectation operator, z is the probability density function; in each iteration, the MBBO algorithm selects the hyperparameter combination that minimizes the objective function f ( θ ), and gradually approaches the global optimal solution through an efficient balance between exploration and utilization strategy.

[0010] S5, calculate the root mean square error, mean absolute error, mean absolute percentage error and determination coefficient based on the test set, and evaluate the prediction performance of the MBBO-LSTM model.

[0011] The prediction performance evaluation system in the present application includes the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and determination coefficient (R 2 ) of the actual value and the predicted value of the power load, which are used to measure the model accuracy and reliability; The actual value and the predicted value of the short-term power load are respectively represented by continuous time sequence vectors and , wherein T denotes the total number of time steps, and the mean value of the actual power load value is denoted as In order to evaluate the prediction performance of the model, the following statistical indicators are set up: The root mean square error RMSE reflects the overall prediction error level and is sensitive to outliers, highlighting the performance of the model in extreme cases, and the formula is as follows: ; In the formula, T denotes the total number of time steps, y t is the actual value of the power load, y t is the actual value of the power load, is the predicted value of the power load; The mean absolute error MAE represents the average deviation between the predicted value and the true value, and the formula is as follows: ; The mean absolute percentage error MAPE has good cross-scale comparability, and the formula is as follows: ; The coefficient of determination R 2 is used to measure the goodness of fit of the predicted value to the true value, and reflects the prediction reliability of the model, and the formula is as follows: ; In the formula, is the mean value of the actual power load value.

[0012] Advantages of the present application: (1) The present application couples the power load historical data with multi-source weather information to construct a time sequence input, making the simulation of the prediction model for the complex operation environment of the power system more three-dimensional and real; in the proposed MBBO-LSTM method, the LSTM model structure and the Bayesian optimization (BO) algorithm are adaptively improved to improve the prediction accuracy, a new PLRDS combined with ESM is introduced to further enrich the theoretical system of power intelligent dispatching and decision support technology.

[0013] (2) The present application selects the hourly power load-weather information of a certain region in northern China in winter and summer in 2023 as the data set for experimental verification; based on the data set, the proposed MBBO-LSTM method is compared with the classical LSTM (Vanilla LSTM) and the current commonly used PSO-LSTM based on particle swarm optimization, GA-LSTM based on genetic algorithm, SCA-LSTM based on sine cosine algorithm and GWO-LSTM based on grey wolf algorithm and other models; the results show that the MBBO-LSTM method realizes significant improvement in training efficiency, stability and load prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is the structural framework diagram of the MBBO-LSTM model of the present application.

[0015] Figure 2 is the implementation program diagram of the short-term power load prediction method based on the MBBO-LSTM model of the present application.

[0016] Figure 3 is the winter typical period power load data set diagram of the present embodiment.

[0017] Figure 4 is the winter typical period power load MBBO-LSTM method prediction result diagram of the present embodiment.

[0018] Figure 5 is the summer typical period power load data set diagram of the present embodiment.

[0019] Figure 6 is the summer typical period power load MBBO-LSTM method prediction result diagram of the present embodiment. DETAILED DESCRIPTION

[0020] The present application will be further described in detail below in combination with specific embodiments.

[0021] As shown in Figure 1 , a short-term power load prediction method based on an MBBO-LSTM model, specifically comprising the following steps: Embodiment 1: The present embodiment is a short-term power load prediction method based on an MBBO-LSTM model, specifically comprising the following steps: S1, collect historical power load data, multi-source weather data and time information, fuse into a multivariate time series, and generate a three-dimensional tensor sample set through sliding window sampling and normalization; divide the sample set into a training set, a validation set and a test set; The present application sets the power load-multisource weather time series obtained by actual observation as , wherein denotes the length of time series, F denotes the number of variables contained in each time step of time series; in t each time step, the time series s is sampled by a sliding window with length w and step X , obtaining input samples ; the corresponding output sample is , which denotes the load value at time t+s; select m continuous samples to , construct a three-dimensional tensor sample set , wherein m denotes the number of samples, that is, the number of continuous sample pairs obtained by sampling the original time series through the sliding window; at the same time, in order to reduce the dimensional difference between different attribute features, the input and output samples are scaled to the interval [0, 1] by using min-max normalization method; When constructing the three-dimensional tensor sample set, the sliding window length s = 24h and the step w = 1h are set, and h denotes hour; the three-dimensional tensor sample set is divided into a training set and a test set according to a ratio of 8:2, and 20% samples are extracted from the training set as a validation set; In this embodiment, a data set composed of hourly power load-multiple source weather data actually measured in winter of 2023 in a certain area in northern China is selected for testing, the time span is from 00:00 on January 1 to 23:00 on February 11, and 1128 continuous load and weather data are included; the mean and variance statistical analysis results of each feature are shown in Table 1: Table 1: Statistical analysis of power load-multiple source weather data in winter As can be seen from the above data, the power load presents a large fluctuation in winter, and the weather variables such as temperature, apparent temperature and visibility also have a significant change range, which reflects that the load change in this period is driven by multiple factors, and higher requirements are put forward for the dynamic response ability and robustness of the model; When constructing the winter data sample set, the sliding window length s = 24h and the step w = 1h are set; the sample set is divided into a training set and a test set according to a ratio of 8:2, and 20% samples are extracted from the training set as a validation set; Figure 3 The hourly change trend of power load in a typical winter period and the division of the training set, the validation set and the test set are respectively shown.

[0022] S2, construct an improved LSTM network model suitable for three-dimensional tensor sample set input; input the training set, and model the multivariate time sequence characteristics through the improved LSTM network model; the improved LSTM network model sequentially comprises a multivariate input layer, a single LSTM layer, a Dropout layer, a full connection layer and a regression output layer; The improved LSTM network model structure of the application is constructed in the following manner: The multivariate input layer receives the training set, the single LSTM layer uses the gating mechanism to establish a memory cell update process, the Dropout layer uses a regularization method to screen data characteristics to reduce the risk of overfitting, the full connection layer integrates data characteristics and maps them to the output space of the prediction task, and the regression layer outputs the characteristic data of the training set after inverse normalization processing, thereby realizing the prediction of the future time change trend of the power load; the working principle of the improved LSTM network model is as follows: The application sets each input sample as a three-dimensional tensor sample set χ with a length of s , wherein each input sample ; in each time step, the improved LSTM network updates its memory cell through the gating mechanism c t , thereby effectively modeling the time dependence in the entire sequence. First, the forgetting gate f t adjusts the retention of the memory of the previous time step, and the formula is as follows: ; Wherein, x t is the current input, h t-1 is the hidden state of the previous moment, W f is the weight of the forgetting gate, b f is the bias term of the forgetting gate. Subsequently, the input gate i t and the candidate cell state u t adjust the influence of the current input on the memory cell, and the formula is as follows: , ; Wherein, W i is the weight of the input gate, W u is the weight of the candidate state, b i is the bias term of the input gate, b uis the bias term of the candidate state, and σ(.) denotes the Sigmoid activation function, tanh(.) denotes the hyperbolic tangent activation function; then, the update of the memory cell state is realized c t , as shown in the following formula: wherein, c t-1 is the memory state of the previous time step, and denotes the element multiplication (Hadamard) operation; Finally, the output gate o t controls the generation of the hidden state h t , as shown in the following formula: wherein, W o is the weight of the output gate, b o is the bias term of the output gate; After the single LSTM layer, the final hidden state h T is input to the Dropout layer, which randomly turns off part of the neurons during the training process to reduce overfitting and feature redundancy while preserving the propagation of time series information; the output after the Dropout layer is processed is input to the fully connected layer and the regression layer, and finally the predicted value of the power load data is output; To verify the effectiveness of the proposed MBBO-LSTM model, a variety of existing LSTM models are selected as comparison benchmarks; the comparison models include: a classic LSTM (Vanilla LSTM) model, a PSO-LSTM model based on particle swarm optimization, a GA-LSTM model based on genetic algorithm, an SCA-LSTM model based on sine cosine algorithm, and a GWO-LSTM model based on grey wolf algorithm; the above models are all based on the unmodified standard LSTM structure, wherein the dimension settings of each layer are as shown in Table 2: Table 2: Network layer dimension settings of comparison models In the comparison model, the LSTM model is trained using the Adam optimizer, with an initial learning rate of 0.01, a maximum training round of 200, and a batch size of 128; to ensure consistent data order, no random shuffling is performed during training, and the training is performed on the CPU; the verification frequency is once every 10 iterations; in addition, the core hyperparameter settings of each optimization algorithm are shown in Table 3: Table 3: Optimization algorithm hyperparameter settings in comparison models All models are run on the same configuration of computing platform; the specific configuration is: the processor is the 13th generation Intel® Core™ i5-13400F, the main frequency is 2.5GHz, and it is equipped with 16.0GB memory.

[0023] S3, the training scheme combining segmented learning rate decay and early ending mechanism is used for dynamic adaptive training of the improved LSTM network model. In the model training process, the learning rate is dynamically adjusted according to the change of the training round number, the segmented learning rate decay (PLRDS) is used to ensure a high learning rate in the early training to accelerate convergence, and then the learning rate is gradually reduced to realize stable optimization. η t , the formula is as follows: ; Among them, η t is the learning rate in the training process, η 0 is the initial learning rate, t e is the current training round number, T d is the preset learning rate decay period, γ is the preset learning rate decay factor, is the floor operation; At the same time, the early ending mechanism (ESM) is used to dynamically monitor the validation loss; every f v validation frequency monitors the validation loss once, if no performance improvement is observed in continuous P times of evaluation, the training process is terminated in advance, thereby effectively preventing overfitting and saving computing resources; In the present application, η 0 the initial learning rate is set to 0.01, f v the validation frequency is set to 10, P and is set to 100 times.

[0024] S4, the improved LSTM network model structure and training hyperparameters are optimized in stages by using the multi-batch Bayesian optimization MBBO algorithm, and the short-term power load prediction method of the MBBO-LSTM model is completed; the hyperparameters include the number of hidden units, the initial learning rate, the learning rate decay factor, the learning rate decay period and the Dropout rate. The present application proposes a MBBO-LSTM model, as Figure 2The MBBO hyperparameter optimization algorithm is used for adjusting the hidden unit number, initial learning rate, learning rate decay factor, learning rate decay period and Dropout rate hyperparameters; the first stage optimizes the LSTM network structure parameters, including the number of hidden units and the initial learning rate, so as to improve the feature representation capability and the model convergence speed; in the application, the number of hidden units is set as [10, 300], and the initial learning rate is set as [1x10 -3 , 0.1]; The second stage adjusts the learning rate decay factor to realize a more stable training process; in the application, the learning rate decay factor is set as [0.001-0.002], and the learning rate decay period is set as [120, 160]; The third stage fine-tunes the Dropout rate to relieve overfitting and improve the model generalization capability, and in the application, the fine-tuned Dropout rate is [0.01, 0.025]; Each stage reconstructs the proxy objective function on the basis of the optimal result of the previous stage, and further searches in the updated hyperparameter space; The application sets the improved LSTM network structure and the hyperparameter set of training as { θ 1, θ 2, ···, θ n} wherein n is the number of hyperparameters, n is 5; and the performance index on the validation set E RMSE is taken as the objective function f ( θ ), and the hyperparameter optimization problem is formalized as follows: ; Wherein, θ * is the optimal hyperparameter combination for minimizing the objective function, and argmin represents the input for taking the minimum value of the objective function; The MBBO algorithm proposed in the application uses a Gaussian process to construct a probability proxy model for the objective function f ( θ ), and estimates the posterior distribution of the optimal hyperparameter combination after observing the sample, so as to predict the function value under any value input: Wherein, μ ( θ ) is the predicted mean, k ( θ , θ ′) is a covariance function, used to describe the correlation and uncertainty between samples; This invention employs Expected Improvement (EI) as the core acquisition function of the MBBO strategy, balancing exploration and utilization when selecting new sampling points. , ; in, f min (θ) The optimal target value observed so far. μ ( θ ) represents the predicted mean of the surrogate model. σ ( θ ) represents the standard deviation of the surrogate model, Φ( z () is the cumulative distribution function of the standard normal distribution. φ ( z Let be the probability density function of the standard normal distribution. E (·) represents the expectation operator. z Let be the probability density function; in each iteration, the MBBO algorithm selects the objective function to minimize. f ( θ By combining hyperparameters and employing an efficient balance exploration and utilization strategy, the global optimal solution is gradually approached. In this embodiment, the MBBO-LSTM method is configured as follows: In the MBBO-LSTM model proposed in this invention, the Adam algorithm is selected as the training optimizer for the improved LSTM network model, the maximum number of training epochs is set to 1000, and the gradient threshold is set to 1 to prevent gradient explosion; the initial learning rate... η The learning rate is determined by the MBBO algorithm and employs a segmented learning rate scheduling strategy: every time the optimal descent cycle determined by the MBBO algorithm is reached, the learning rate is adjusted. η Multiply by the optimal learning rate reduction factor obtained by the MBBO algorithm; clean the data once in each round of training and perform validation once after every 10 iterations. If the validation set loss does not improve in 100 consecutive iterations, stop training early to prevent overfitting; the parameter settings of the MBBO method are shown in Table 4: Table 4: Hyperparameter settings for the MBBO method To verify the effectiveness of the training scheme combining PLRDS and ESM proposed in this invention, the training process of the LSTM models of all algorithms was monitored using training loss and validation loss. Under winter data conditions, the training losses of MBBO-LSTM and the comparison algorithms in the first 240 iterations are shown in Table 5. Table 5: Comparison of training loss changes for different models in the first 240 iterations on winter data. As can be seen from Table 5, the proposed MBBO-LSTM model has the lowest overall validation loss during the entire training process, showing better convergence and prediction stability; especially at the 240th iteration, its validation loss is 0.0483, which is better than all the comparative models, indicating that the model has stronger fitting ability and generalization performance in the winter load prediction task; SCA-LSTM also performs well, with lower validation loss at multiple iteration points, but overall it is slightly inferior to MBBO-LSTM; In this embodiment, the validation loss of each model is calculated on multiple independently divided validation subsets to more accurately reflect its generalization ability; the validation loss results of each model on different validation sets are shown in Table 6: Table 6: Validation loss of different models on different validation sets of winter data As can be seen from Table 6, the proposed MBBO-LSTM model achieves lower validation loss on most validation subsets; it especially reaches the lowest value on validation subset 6, showing its superior generalization ability and strong stability; in contrast, Van-LSTM has large fluctuations in loss on multiple validation sets, indicating that it is sensitive to hyperparameter settings and lacks robustness; other optimization models based on evolutionary algorithms, such as PSO-LSTM, GA-LSTM, SCA-LSTM and GWO-LSTM, although they achieve relatively good results in some validation sets, but the overall validation loss values and consistency are slightly inferior to MBBO-LSTM; this further verifies the effectiveness and advantages of the optimization strategy proposed in this study in improving the efficiency of hyperparameter selection and model performance.

[0025] S5, calculate the root mean square error, mean absolute error, mean absolute percentage error and determination coefficient based on the test set to evaluate the prediction performance of the MBBO-LSTM model.

[0026] The prediction performance evaluation system in the present application includes prediction performance indicators such as root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and determination coefficient (R 2 ) of actual and predicted values of power load, which are used to measure model accuracy and reliability; In the present application, the actual and predicted values of short-term power load are represented as continuous time series vectors as follows: and wherein TThis represents the total number of time steps; the average value of the actual power load is denoted as... To evaluate the predictive performance of the model, the following statistical indicators are established: The root mean square error (RMSE) reflects the overall prediction error level and is sensitive to outliers, highlighting the model's performance in extreme cases. The formula is shown below: ; In the formula, T represents the total number of time steps. y t This represents the actual value of the power load. y t This represents the actual value of the power load. This is the predicted value of the power load; Mean Absolute Error (MAE) represents the average deviation between the predicted and actual values, and the formula is shown below: ; Mean Absolute Percentage Error (MAPE) has good cross-dimensional comparability, as shown in the following formula: ; Coefficient of determination R 2 The formula used to measure the goodness of fit between predicted and actual values, reflecting the reliability of model predictions, is shown below: ; In the formula, This is the average of the actual power load values; Among the above evaluation metrics, the smaller the RMSE, MAE, and MAPE values, the closer R² is to 1, indicating that the prediction model performs better. The prediction performance of MBBO-LSTM and the contrastive model on a test set during typical winter periods in this embodiment is as follows: Figure 4 As shown; by Figure 4 It can be clearly seen that the MBBO-LSTM prediction results are closest to the actual load values, and the overall trend is highly consistent. In contrast, the dashed prediction curves corresponding to the other comparison algorithms show significant deviations from the actual values, indicating that their fitting accuracy and trend tracking ability are relatively weak. To comprehensively evaluate the performance of each model in the winter load forecasting task, a comparative analysis of their prediction error indices is further conducted, and the results are shown in Table 7. Table 7: Comparison of prediction performance metrics of different models on the winter data test set comprehensive Figure 4 As shown in Table 7, MBBO-LSTM achieves the best results in all prediction performance metrics; its RMSE, MAE, MAPE, and R... 2The calculation results of the MBBO-LSTM are significantly better than those of the other comparative models, showing extremely strong prediction accuracy and fitting ability; in contrast, the SCA-LSTM performs sub-optimally in some indicators, especially in MAPE and MAE, which are close to those of the MBBO-LSTM, but its R 2 is only 0.7666, indicating that its overall fitting quality is slightly inferior; although the RMSE and MAE of the Van-LSTM are slightly better than those of the models optimized by the evolutionary algorithm, its MAPE is as high as 10.8546, and its R 2 is only 0.5079, indicating that it is unstable in responding to volatile loads and is prone to systematic bias; the PSO-LSTM, GA-LSTM, and GWO-LSTM all perform poorly in all indicators, especially the GA-LSTM, whose R 2 is only 0.1074, almost without explanatory power, and the prediction result deviates greatly from the true value; In summary, the MBBO-LSTM performs best in winter load prediction in terms of robustness and accuracy, verifying the effectiveness of the proposed optimization strategy in improving the model's generalization ability and reducing prediction errors; In combination with the principle module design of S1 to S5, the MBBO-LSTM method is used for short-term electric power load prediction in a complex environment of multi-source information fusion, and its calculation process can be divided into three stages as shown in Figure 1 : data preparation, model construction, and performance evaluation; first, the multi-variable time series of the collected multi-source meteorological information and historical electric power load data are preprocessed to ensure that the data structure matches the modeling requirements and is divided into a training set, a validation set, and a test set; then, the improved LSTM network is constructed to receive input data coupled with multiple features, and the MBBO algorithm is used to adjust the network structure and obtain the optimal hyperparameters to start the training process; finally, the predicted results output by the model are de-normalized, and the evaluation index system is used to select the optimal prediction results of the electric power load.

[0027] Example 2: The structure of the steps is the same as that of Example 1, except that this example is a simulation experiment and result analysis of short-term electric power load prediction based on a real summer data set; This example selects a data set composed of hourly electric power load-meteorological information actually measured in a certain region in northern China in the summer of 2023 for testing, with a time span of 00:00 on June 1 to 23:00 on July 12, containing 1128 consecutive load and meteorological data; the mean and variance statistical analysis results of each feature are shown in Table 8: Table 8: Statistical analysis of typical period electric power load-meteorological information features in summer As can be seen from the above data, the load demand and each meteorological index in summer are generally high, reflecting the significant influence of seasonal changes on electricity demand and environmental conditions; In the process of constructing the summer data sample set, the sliding window length s = 24h and the step length w = 1h are set; the sample set is divided into a training set and a test set according to a ratio of 8:2, and 20% of the samples in the training set are extracted as a validation set; Figure 5 The hour change trend of the typical summer period power load and the division of the training set, the validation set and the test set are respectively shown; In order to verify the effectiveness of the training scheme combining the PLRDS and the ESM proposed in the application, the training loss and the validation loss are used to monitor the training process of the LSTM model of all algorithms; under the summer data condition, the training loss of the MBBO-LSTM and the comparative algorithms in the first 240 iterations is shown in Table 9: Table 9: Comparison of training loss changes of different models in the first 240 iterations under summer data As can be seen from Table 9, the MBBO-LSTM maintains a low training loss in most training stages; the model achieves the lowest value at the 240th iteration, reflecting its excellent convergence performance; in the whole training process, the loss change curve of the MBBO-LSTM presents a more smooth downward trend, indicating that its training process is more stable; in comparison, although the PSO-LSTM and the GA-LSTM perform similarly in some stages, there are still some fluctuations at multiple iteration points; especially the GWO-LSTM, its training loss fluctuates significantly in the training process, and even rebounds in the later stage, indicating that it is not adaptive to the summer high temperature load condition; in addition, the Van-LSTM has a high training loss in the initial stage, although it temporarily decreases near the 160th iteration, but rises again in the later stage, and the overall stability and convergence ability are not as good as the MBBO-LSTM; In this embodiment, the validation loss of each model is calculated on multiple independently divided validation subsets to more accurately reflect its generalization ability; the validation loss results of each model on different validation sets are shown in Table 10: Table 10: Validation loss of different models on different validation sets of summer data As can be seen from Table 10, the MBBO-LSTM has five validation losses lower than 0.2 in the six validation subsets, especially in the validation subset 2 to the validation subset 6, which achieves the lowest or the second lowest value, showing significant generalization ability and robustness; The prediction performance of this embodiment on the summer typical period test set is shown in Table 11: Figure 6 Table 11: Prediction performance of different models on the summer typical period test set Figure 6It can be seen directly that the MBBO-LSTM prediction curve is highly consistent with the actual load in overall trend and local fluctuation; especially in the stage of rapid load rise, MBBO-LSTM can accurately capture the peak change, showing good dynamic response ability; in contrast, the prediction curves of the remaining models generally have a certain degree of lag or high-frequency oscillation distortion; especially during the period of rapid load change, the prediction deviation of the comparative algorithm is more obvious; In order to comprehensively evaluate the performance of each model in summer load prediction task, further comparative analysis of the prediction error indicators is carried out, and the results are shown in Table 11: Table 11: Comparison of prediction performance indicators of different models on summer data test set Comprehensive Figure 6 As can be seen from Table 11, MBBO-LSTM also shows the best prediction performance in summer typical period; this model is better than other comparative models in four key indicators of RMSE, MAE, MAPE and R², which fully verifies its robustness and stability under high temperature and high fluctuation load conditions; it is worth noting that SCA-LSTM performs close to MBBO-LSTM in MAE and MAPE indicators, indicating that it has certain advantages in overall trend fitting; however, its R² value is 0.8111, which is still lower than that of MBBO-LSTM, 0.8374, indicating that its fitting stability is still insufficient; Van-LSTM performs in the middle, although its RMSE and MAE are slightly better than those of some optimization models, but its MAPE reaches 6.9186%, and R² is only 0.6707, showing that its response to rapidly changing load is relatively slow and easy to have systematic deviation; PSO-LSTM and GA-LSTM achieve certain results in some indicators, but the overall error is still higher than that of MBBO-LSTM, indicating that the optimization effect is limited; the error indicators of GWO-LSTM are all at a high level, especially the MAPE reaches 7.1024%, reflecting that its fitting ability in summer high disturbance scene is relatively weak; In summary, MBBO-LSTM shows stronger generalization ability and prediction accuracy when dealing with high nonlinearity and high fluctuation characteristics in summer load prediction. This further verifies the robustness and adaptability of the proposed optimization strategy under multi-seasonal conditions.

[0028] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application; therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A short-term power load forecasting method based on an MBBO-LSTM model, characterized in that, The method comprises the following steps: S1, collecting historical power load data, multi-source weather data and time information, fusing into a multivariate time series, and generating a three-dimensional tensor sample set through sliding window sampling and normalization; dividing the sample set into a training set, a validation set and a test set; S2, constructing an improved LSTM network model suitable for the input of the three-dimensional tensor sample set; inputting the training set, and modeling the multivariate time series characteristics through the improved LSTM network model; The improved LSTM network model comprises, in sequence, a multivariate input layer, a single LSTM layer, a Dropout layer, a full connection layer and a regression output layer; S3, using a training scheme combining segmented learning rate decay and early end mechanism to dynamically adaptively train the improved LSTM network model; S4, using an MBBO algorithm to stage-optimize the structure and training hyperparameters of the improved LSTM network model, and completing a short-term power load prediction task based on the MBBO-LSTM model; the hyperparameters include the number of hidden units, the initial learning rate, the learning rate decay factor, the learning rate decay period and the Dropout rate; S5, calculating the root mean square error, the mean absolute error, the mean absolute percentage error and the determination coefficient based on the test set to evaluate the prediction performance of the MBBO-LSTM model. 2.The short-term power load forecasting method based on the MBBO-LSTM model according to claim 1, wherein, The historical power load data, multi-source weather data and time information are collected, fused into a multivariate time series, and a three-dimensional tensor sample set is generated through sliding window sampling and normalization, wherein the multi-source weather data includes eight weather variables of temperature, humidity, wind speed, air pressure, precipitation, visibility, water vapor pressure and apparent temperature. 3.The short-term power load forecasting method based on the MBBO-LSTM model according to claim 1, characterized in that, The collection history power load data, multi-source weather data and time information are fused into a multivariate time series, and a three-dimensional tensor sample set is generated through sliding window sampling and normalization. The way to generate the three-dimensional tensor sample set through sliding window sampling and normalization is: setting the sliding window length s = 24h, the step w = 1h, h represents hour, and adopting min- max The normalization operation is performed. 4.The short-term power load forecasting method based on the MBBO-LSTM model of claim 1, wherein, The three-dimensional tensor sample set is divided into a training set, a validation set and a test set in the following manner: the three-dimensional tensor sample set is divided into a training set and a test set according to a preset ratio, and 20% of the samples in the training set are extracted as a validation set. 5.The short-term power load forecasting method based on the MBBO-LSTM model according to claim 1, wherein, In the improved LSTM network model in S2: the multivariate input layer receives the training set, establishes a memory unit update process through the single LSTM layer and using the gating mechanism, the Dropout layer uses the regularization method to screen data features to reduce the risk of overfitting, the full connection layer integrates data features and maps them to the output space of the prediction task, and the regression layer outputs the feature data of the training set after inverse normalization, thereby realizing the prediction of the future time variation trend of the power load. 6.The short-term power load forecasting method based on the MBBO-LSTM model according to claim 1, wherein, The improved LSTM network model is dynamically adaptively trained using a training scheme combining segmented learning rate decay and early end mechanism; the formula of the segmented learning rate decay is: ； wherein, η t is a learning rate during training, η 0 is an initial learning rate, t e is a current training epoch, T d is a preset learning rate decay period, γ is a preset learning rate decay factor, is a floor operation; the early stopping mechanism is used to prevent overfitting.

7. The short-term power load forecasting method based on the MBBO-LSTM model according to claim 1, characterized in that, The MBBO algorithm stage-optimizes the structure and training hyperparameters of the improved LSTM network model in the following three stages: The first stage optimizes the LSTM network structure parameters, including the number of hidden units and the initial learning rate; The second stage adjusts the learning rate decay factor and the learning rate decay period in the training scheme; The third stage adjusts the Dropout rate; The second stage and the third stage reconstruct the agent objective function on the basis of the optimal result of the previous stage, and search in the updated hyperparameter space; the reconstruction of the agent objective function is as follows: taking the performance index on the verification set as the agent objective function E RMSE As the agent objective function f θ , a Gaussian process is adopted to construct a probability agent model for the agent objective function f θ The reconstruction of the agent objective function balances the exploration and the utilization when selecting a new sampling point by taking the expected improvement as the core acquisition function of the MBBO strategy, and the formula of the expected improvement α EI ( θ ) is as follows:​​ ； ； wherein, f min ( θ ) is the current observed optimal target value, μ ( θ ) is the predicted mean of the surrogate model, σ ( θ ) is the standard deviation of the surrogate model, Φ ( z ) is the cumulative distribution function of the standard normal distribution, φ ( z ) is the probability density function of the standard normal distribution, E(·) is the expectation operator, z is the probability density function; the hyperparameter combination of the objective function f ( θ ) is selected by the MBBO algorithm to obtain the global optimal solution. 8.The short-term power load forecasting method based on the MBBO-LSTM model of claim 1, wherein, The root mean square error, the mean absolute error, the mean absolute percentage error and the determination coefficient are calculated based on the test set, and the expressions of the root mean square error, the mean absolute error, the mean absolute percentage error and the determination coefficient in the evaluation of the prediction performance of the MBBO-LSTM model are as follows: The root mean square error RMSE reflects the overall prediction error level and is sensitive to outliers, highlighting the performance of the model in extreme cases, and the formula is as follows: ； wherein T is the total number of time steps, y t is the actual value of the power load, is the predicted value of the power load, i is the number of samples; The mean absolute error MAE represents the average deviation between the predicted value and the true value, and the formula is as follows: ； The mean absolute percentage error MAPE has good cross-calculation comparison, and the formula is as follows: ； Coefficient of determination R 2 For measuring the goodness of fit of the predicted value to the true value, reflecting the credibility of the model prediction, the formula is as follows: ； In the formula, Let the average of the actual power load values be denoted by

Images