VMD-BO-LSTM-based multi-feature power load prediction method

Abstract

The invention belongs to the technical field of power load prediction, and particularly relates to a VMD-BO-LSTM-based multi-feature power load prediction method. The method comprises the following steps: acquiring daily load data and related characteristic factor data of a power system, and preprocessing the daily load data and the related characteristic factor data; dividing the preprocessed data into a training set, a verification set and a test set; variational mode decomposition is carried out on the daily load data, and K intrinsic mode functions are obtained; establishing a long short-term memory network model for each intrinsic mode function, and training the long short-term memory network model through the training set; optimizing hyper-parameters of the long-short-term memory network model through the verification set and a Bayesian optimization algorithm; and evaluating the optimized long and short-term memory network model through the test set. The method shows good prediction performance, can effectively process the non-linear and non-stationary characteristics of the power load data, and significantly improves the prediction precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load prediction, and in particular relates to a multi-feature power load prediction method based on VMD-BO-LSTM. Background Art

[0002] As the scale of power systems continues to expand and the complexity continues to increase, accurate prediction of power system load is of great significance to ensure the safe and stable operation of power systems. At present, common power load forecasting methods mainly have the following problems:

[0003] (1) The nonlinearity and non-stationarity of the data affect the prediction accuracy.

[0004] (2) The overfitting problem of traditional deep learning models in load data processing leads to poor generalization ability of the model.

[0005] (3) It relies heavily on input feature selection and model hyperparameter optimization, which increases the complexity of the prediction model.

[0006] At present, commonly used methods include support vector machines, random forests, neural networks and their extended models. However, these methods have limitations when processing complex load data, and traditional hyperparameter tuning methods are difficult to effectively improve model performance. At the same time, the comprehensive utilization of multidimensional characteristic factors in existing research is insufficient, which affects the practicality of the prediction model. Summary of the invention

[0007] The present invention provides a multi-feature power load forecasting method based on VMD-BO-LSTM, which applies variational mode decomposition (VMD) and Bayesian optimization (BO) technology to a long short-term memory network (LSTM), and proposes an efficient and accurate multi-feature power load forecasting method.

[0008] The technical solution of the present invention is as follows: a multi-feature power load forecasting method based on VMD-BO-LSTM, comprising:

[0009] S10: acquiring daily load data and relevant characteristic factor data of the power system, and preprocessing the daily load data and relevant characteristic factor data;

[0010] S20: Divide the preprocessed data into a training set, a validation set and a test set;

[0011] S30: performing variational mode decomposition on the daily load data to obtain K intrinsic mode functions;

[0012] S40: establishing a long short-term memory network model for each intrinsic mode function, and training the long short-term memory network model through a training set;

[0013] S50: Optimize the hyperparameters of the long short-term memory network model using the validation set and Bayesian optimization algorithm;

[0014] S60: Evaluate the optimized LSTM network model using the test set.

[0015] Furthermore, the step S10 includes:

[0016] Performing data cleaning on the daily load data and related characteristic factor data, and deleting abnormal data and missing data;

[0017] The relevant characteristic factor data were screened through Spearman correlation analysis and heat map analysis to select the core characteristic factor data.

[0018] Furthermore, in step S30:

[0019] The objective function of the variational mode decomposition is:

[0020]

[0021] In the formula, u k (t) is the kth eigenmode function; ω k is the center frequency of the kth intrinsic mode function; * denotes the convolution operation; δ(t) is the Dirac function; j is the imaginary unit, is the time derivative.

[0022] Furthermore, in step S40:

[0023] Each intrinsic mode function builds a long short-term memory network model, including an input layer, two LSTM layers, a Dropout layer, a fully connected layer, and a regression output layer;

[0024] Set the initial values ​​of the hyperparameters of the LSTM network model.

[0025] Furthermore, the hyperparameters are the hidden units of the first layer of the neural network, the hidden units of the second layer of the neural network, the Dropout ratio and the learning rate.

[0026] Furthermore, the step S50 includes:

[0027] Set the hyperparameter search space for the LSTM model;

[0028] Select the optimization objective function;

[0029] Run the Bayesian optimization algorithm to iterate the hyperparameters of the LSTM network model;

[0030] The optimal hyperparameters are selected and configured in the LSTM network model.

[0031] Furthermore, the optimization objective function is the mean square error of the validation set.

[0032] Furthermore, in step S50, the formula of the Bayesian optimization algorithm is:

[0033] EI(x)=Ε[max(0,f(x)-f(x * ))]

[0034] Among them, f(x) is the optimization objective function, EI(x) is the function to be improved, and x * is the location of the optimal sampling point, f(x * ) is the current optimal solution.

[0035] Furthermore, the step S60 includes:

[0036] Input the core feature factor data of the test set into the trained LSTM network model;

[0037] Reconstruct the K intrinsic mode function values ​​output by the long short-term memory network model to obtain the predicted load;

[0038] The indicator evaluation is carried out based on the predicted load and the daily load data of the test set.

[0039] Furthermore, the specific indicators of the indicator evaluation include: MAE, RMSE, MAPE, R 2 and NSE.

[0040] Beneficial effects of the present invention:

[0041] (1) Improving prediction accuracy: VMD decomposition is used to eliminate the non-stationarity of load data, enhance the learning ability of the LSTM model, and use BO to optimize the LSTM hyperparameters, thereby improving the overall prediction performance of the model.

[0042] (2) Automated hyperparameter optimization: The Bayesian optimization method significantly reduces the hyperparameter tuning time and improves the generalization ability and stability of the model.

[0043] (3) Effectiveness of multi-feature input: Make full use of multi-dimensional feature information to make the model more interpretable and adaptable in capturing load change characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the present invention.

[0045] Figure 2 It is a structural diagram of the framework of the present invention. DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. The described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0047] In one technical solution of the present invention, Figure 1 This is a flowchart provided according to the specific process of a multi-feature power load forecasting method based on VMD-BO-LSTM in the technical solution, such as Figure 1 As shown, the present invention specifically includes:

[0048] S10: Acquire daily load data and relevant characteristic factor data of the power system, and preprocess the daily load data and relevant characteristic factor data.

[0049] Collect the daily maximum load data of the power system and relevant characteristic factor data related to the daily maximum power load, such as meteorological (temperature, humidity, wind speed, etc.), economic (GDP), population and holiday information.

[0050] The daily load data and related characteristic factor data are cleaned, abnormal data and missing data are deleted, and necessary data conversion is performed.

[0051] The relevant characteristic factor data are screened through Spearman correlation analysis and heat map analysis to screen out the core characteristic factor data. The most relevant core characteristic factors are screened out using methods such as Spearman correlation analysis and heat map analysis.

[0052] S20: Divide the preprocessed data into a training set, a validation set, and a test set.

[0053] The processed data set is divided into training set, validation set and test set. Specifically, the training set is used for model training, the validation set is used to adjust hyperparameters and preliminarily evaluate model performance during model training, and the test set is used to finally evaluate the generalization ability of the model.

[0054] S30: performing variational mode decomposition on the daily load data to obtain K intrinsic mode functions.

[0055] The variational mode decomposition (VMD) is specifically performed on the original load data to obtain multiple intrinsic mode functions (IMFs). The objective function of the variational mode decomposition is:

[0056]

[0057] In the formula, uk (t) is the kth eigenmode function; ω k is the center frequency of the kth intrinsic mode function; * denotes the convolution operation; δ(t) is the Dirac function; j is the imaginary unit, is the time derivative.

[0058] Set the algorithm parameters in the above objective function, such as the number of decomposition modes K, etc.

[0059] The VMD algorithm works as follows:

[0060] The first step is initialization. Initialize the modal function u of each mode k (t), center frequency ω k and the Lagrange multiplier λ(t).

[0061] The second step is alternating optimization. For each mode u k (t), keeping other modes and frequencies unchanged, by adjusting the objective function u k The derivative of (t) is zero. Solve to optimize the current modal function u k (t). Update the modal center frequency ω k , update the frequency according to the frequency domain energy distribution of the current modal function. Update the Lagrange multiplier λ(t), and update the multiplier by the gradient ascent method. The formula of the gradient ascent method is as follows:

[0062]

[0063] The third step is cyclic update. The modal function, center frequency and Lagrange multiplier are continuously updated alternately until the convergence condition formula is met, that is, the change of the modal function is less than the preset threshold ξ, as shown in the formula below, or the maximum number of iterations is reached.

[0064]

[0065] The fourth step is final decomposition. After several rounds of iterations, the decomposed modal functions u are output k (t), and the corresponding center frequency ω k .

[0066] S40: Establish a long short-term memory network model for each intrinsic mode function, and train the long short-term memory network model through the training set.

[0067] In step S40:

[0068] Each intrinsic mode function builds a long short-term memory network model, including an input layer, two LSTM layers, a Dropout layer, a fully connected layer, and a regression output layer;

[0069] Set the initial values ​​of the hyperparameters of the long short-term memory network model. Specifically, the hyperparameters are the hidden units of the first layer of the neural network, the hidden units of the second layer of the neural network, the dropout ratio, and the learning rate.

[0070] Long short-term memory network (LSTM) is a special recurrent neural network (RNN). By introducing memory units and gating mechanisms, it can effectively solve the gradient vanishing problem of traditional RNN when processing long sequences. For each IMF component, an LSTM network structure consisting of an input layer, LSTM layer, Dropout layer, and output layer is established, and the hyperparameter configuration is initialized.

[0071] x t is the input quantity at the current moment, C t , C t-1 Storage unit information for the current moment and the previous moment; The storage unit value adjusted for information; h t 、h t-1 are the output vectors of the hidden layer at the current moment and the previous moment, σ is the sigmoid function, and tanh is the activation function; i t , o t 、f t They are input gate, output gate and forget gate respectively.

[0072] The core of LSTM lies in its unique memory unit and three gating mechanisms, input gate, forget gate and output gate. These components work together to enable LSTM to selectively store, update and output information. Its working principle can be summarized in the following steps:

[0073] The specific principles are as follows:

[0074] The first step is to input data. LSTM receives input sequence data.

[0075] The second step is to update the memory unit. t Determine the current state to determine the information that needs to be discarded from the cell state. The specific formula is:

[0076] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0077] Where W f and b f are weight and bias parameters respectively, and σ is the activation function.

[0078] Step 3: Update the state. Input gate i t Determine the information that needs to be updated. The specific formula is:

[0079]

[0080] In the formula, is a candidate state, and the specific formula is:

[0081]

[0082] Step 4: Output calculation. Output gate o t Determines the hidden state h at the current moment t The specific formula for the output part is:

[0083] h t =o t *tanh(C t )

[0084] In the formula, h t Should be the output of the LSTM at the current time step, used for prediction or as input for the next time step.

[0085] The fifth step is loop iteration. This process is applied to each time step in the time series through the loop structure, and finally the sequence information is integrated into the hidden state.

[0086] Step 6: Output prediction. LSTM makes output predictions based on the hidden state. The output can be the predicted value of each time step or the final state of the entire sequence, which is used for subsequent tasks such as time series prediction.

[0087] S50: Optimize the hyperparameters of the long short-term memory network model through the validation set and Bayesian optimization algorithm.

[0088] Set the hyperparameter search space for the Long Short-Term Memory Network model.

[0089] An optimization objective function is selected, wherein the optimization objective function may be the mean square error of the validation set.

[0090] Run the Bayesian optimization algorithm to iterate the hyperparameters of the LSTM network model.

[0091] The optimal hyperparameters are selected and configured in the LSTM network model.

[0092] Specifically, the formula of the Bayesian optimization algorithm is:

[0093] EI(x)=Ε[max(0,f(x)-f(x * ))]

[0094] Among them, f(x) is the optimization objective function, EI(x) is the function to be improved, and x * is the location of the optimal sampling point, f(x* ) is the current optimal solution.

[0095] Bayesian optimization is an efficient global optimization method, especially suitable for black box function optimization problems with high computational cost. In deep learning, BO is mainly used to automatically adjust hyperparameters to improve model performance. The validation set objective function is set to mean square error (RMSE), and the BO algorithm is used to automatically search for the hyperparameter combination of the LSTM model to optimize the number of hidden units, learning rate, and Dropout ratio of the neural network.

[0096] The core idea of ​​BO is to model the objective function through a proxy model (such as Gaussian process) to reduce the number of actual function evaluations. Its workflow is as follows:

[0097] The first step is initial sample collection. Randomly select several input points {x i}, evaluate the objective function f(x) and obtain the initial sample data {(x i ,f(x i ))}.

[0098] The second step is to build a surrogate model. Using the initial sample data, a surrogate model is built to fit the target function. The commonly used surrogate model is the Gaussian process (GP), as shown below, which can predict the mean μ(x) and variance σ(x) of the target function by a given input x.

[0099] f(x)~N(μ(x),σ 2 (x)

[0100] Where μ(x) is the mean function and σ(x) is the prediction variance.

[0101] The third step is to select a function to be improved (such as expected improvement EI) and maximize the acquisition function on the proxy model to determine the next input point x to be sampled. next , the expected improved formula is:

[0102] EI(x)=Ε[max(0,f(x)-f(x * ))]

[0103] In the formula, x * is the location of the optimal sampling point, f(x * ) is the current optimal solution. During the calculation process, the EI function uses the mean and variance of the GP prediction to calculate the improved expected value.

[0104] The fourth step is to sample and update the model. At the newly selected input point x next On the true objective function f(x next ) is evaluated and the new data point (x next ,f(xnext ))Add to the sample set and update the proxy model.

[0105] Step 5: Iterate. Iterate optimization: Repeat steps 3-4 until the stop condition is met.

[0106] The advantage of BO is that it can find the global optimal solution in a smaller number of iterations, which is particularly suitable for hyperparameter optimization of deep learning models such as LSTM.

[0107] S60: Evaluate the optimized LSTM network model using the test set.

[0108] Input the core feature factor data of the test set into the trained LSTM network model;

[0109] Reconstruct the K intrinsic mode function values ​​output by the long short-term memory network model to obtain the predicted load;

[0110] The indicator evaluation is carried out based on the predicted load and the daily load data of the test set.

[0111] Specifically, we calculate MAE, RMSE, MAPE, R 2 And NSE and other evaluation indicators, and then compare the performance with the traditional model.

[0112] This paper takes the daily power load forecast in Wuxi, Jiangsu as an example to conduct empirical analysis and performance evaluation of the proposed VMD-BO-LSTM model.

[0113] The present invention uses the daily maximum load data of Wuxi, Jiangsu from June 1, 2022 to August 31, 2024, with a total of 767 records. Average temperature, minimum temperature, maximum temperature, humidity, average wind speed, total precipitation, holiday judgment, GDP value, and population are selected as core characteristic factors.

[0114] The data preprocessing steps are as follows: First, remove abnormal data, such as abnormal load data during orderly electricity consumption; second, encode the working day characteristics: Monday to Friday is marked as 1, Saturday is marked as 0.5, and Sunday and holidays are marked as 0; third, process the GDP data: use the quarterly GDP value.

[0115] The Spearman correlation analysis method was used to screen the characteristic factors.

[0116] Based on the results of correlation analysis, average temperature, minimum temperature, maximum temperature, working day judgment, and GDP are selected as core characteristic factors. Considering the small sample size, the cross-validation method is used to improve the accuracy. In this example, the ratio of training set, validation set, and test set is 70:15:15.

[0117] In order to retain the important characteristics of the daily maximum power load and ensure the consistency of the model input, the original data is subjected to variational mode decomposition (VMD) before normalization. In order to ensure the resolution of the balanced signal, improve the frequency structure accuracy of the noise, reduce frequency band aliasing, set the center frequency bandwidth parameters and the number of modes, and the original daily maximum power load can be decomposed into 12 internal mode functions (IMFs).

[0118] In order to improve the learning and generalization capabilities of the model, the basic structure of this LSTM network is set as follows, including an input layer, two LSTM layers, a drop layer in the middle, a fully connected layer, and a regression output layer.

[0119] The initial model parameters are set as follows: the maximum number of training rounds is 200, the training mini-batch size is 32, and the gradient clipping rate is 1.

[0120] The hyperparameters to be optimized for this neural network are set as follows: the initial value of the hidden unit of the first layer of the neural network is set to 128. The initial value of the hidden unit of the second layer of the neural network is set to 64, the initial value of the Dropout ratio is set to 0.2, and the initial value of the learning rate is set to 0.001.

[0121] The mean square error (RMSE) of the validation set was selected as the objective function, and the hyperparameters of LSTM were optimized using Bayesian optimization (BO), and the training set data was iterated.

[0122] During multiple iterations, the optimal feasible point of the mean square error (RMSE) of the model validation set was obtained, and the optimal hyperparameters of LSTM were as follows: the hidden units of the first layer of the neural network were 164, the hidden units of the second layer of the neural network were 26, the Dropout ratio was 0.00192, and the learning rate was 0.00464.

[0123] The 12 IMF components obtained by variational mode decomposition (VMD) of the training set are normalized as output values, and the core characteristic factors are normalized as input values. They are substituted into the aforementioned hyperparameter optimized BO-LSTM neural network for fitting. Finally, the predicted 12 IMF components are obtained by denormalization. The above components are reconstructed to obtain the maximum daily electricity load of the training set, and the VMD-BO-LSTM model can be iteratively obtained.

[0124] The test set data is input into the trained VMD-BO-LSTM model for prediction, and the prediction results are compared with those of the traditional LSTM and VMD-LSTM models. The VMD-BO-LSTM model proposed in the present invention outperforms the traditional LSTM and VMD-LSTM models in all evaluation indicators. In particular, in terms of the two key indicators of MAPE and R2, the VMD-BO-LSTM model achieved excellent results of 1.85% and 0.9693, respectively, which reduced the error value by 10.36% and improved the fit by 0.6337 compared with the traditional LSTM model.

[0125] The VMD-BO-LSTM model of the present invention shows good prediction performance, can effectively handle the nonlinear and non-stationary characteristics of power load data, and significantly improves the prediction accuracy. Compared with traditional methods, this model has good performance in MAPE and R 2 Significant improvements have been made in key indicators such as load forecasting, providing new solutions for accurate load forecasting of power systems.

[0126] The introduction of multidimensional characteristic factors greatly enhances the interpretability and adaptability of the model. By comprehensively considering meteorological, economic and social factors, the model can more comprehensively capture various external factors that affect load changes, thereby providing more accurate and reliable prediction results.

[0127] The Bayesian optimization algorithm plays a key role in the hyperparameter tuning of the LSTM model, effectively solving the problem that traditional manual parameter tuning is time-consuming and unstable. This adaptive optimization method not only improves model performance, but also enhances the generalization ability of the model.

[0128] The application of variational mode decomposition technology effectively reduces the complexity of the original load data and provides clearer and more effective input features for the subsequent deep learning model, thereby improving the overall prediction accuracy.

[0129] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-feature power load forecasting method based on VMD-BO-LSTM, characterized in that: include: S10: acquiring daily load data and relevant characteristic factor data of the power system, and preprocessing the daily load data and relevant characteristic factor data; S20: Divide the preprocessed data into a training set, a validation set and a test set; S30: performing variational mode decomposition on the daily load data to obtain K intrinsic mode functions; S40: establishing a long short-term memory network model for each intrinsic mode function, and training the long short-term memory network model through a training set; S50: Optimize the hyperparameters of the long short-term memory network model using the validation set and Bayesian optimization algorithm; S60: Evaluate the optimized LSTM network model using the test set.

2. The multi-feature power load forecasting method based on VMD-BO-LSTM as claimed in claim 1, characterized in that: The step S10 comprises: Performing data cleaning on the daily load data and related characteristic factor data, and deleting abnormal data and missing data; The relevant characteristic factor data were screened through Spearman correlation analysis and heat map analysis to select the core characteristic factor data.

3. The multi-feature power load forecasting method based on VMD-BO-LSTM as claimed in claim 1, characterized in that: In step S30: The objective function of the variational mode decomposition is: In the formula, u k (t) is the kth eigenmode function; ω k is the center frequency of the kth intrinsic mode function; * denotes the convolution operation; δ(t) is the Dirac function; j is the imaginary unit, is the time derivative.

4. The multi-feature power load forecasting method based on VMD-BO-LSTM as claimed in claim 1, characterized in that: In step S40: Each intrinsic mode function builds a long short-term memory network model, including an input layer, two LSTM layers, a Dropout layer, a fully connected layer, and a regression output layer; Set the initial values ​​of the hyperparameters of the LSTM network model.

5. The multi-feature power load forecasting method based on VMD-BO-LSTM as claimed in claim 4, characterized in that: The hyperparameters are the hidden units of the first layer of the neural network, the hidden units of the second layer of the neural network, the Dropout ratio and the learning rate.

6. The multi-feature power load forecasting method based on VMD-BO-LSTM according to claim 1, characterized in that: The step S50 comprises: Set the hyperparameter search space for the LSTM model; Select the optimization objective function; Run the Bayesian optimization algorithm to iterate the hyperparameters of the LSTM network model; The optimal hyperparameters are selected and configured in the LSTM network model.

7. The multi-feature power load forecasting method based on VMD-BO-LSTM according to claim 6, characterized in that: The optimization objective function is the mean square error of the validation set.

8. The multi-feature power load forecasting method based on VMD-BO-LSTM as claimed in claim 6, characterized in that: In step S50, the formula of the Bayesian optimization algorithm is: EI(x)=Ε[max(0,f(x)-f(x * ))] Among them, f(x) is the optimization objective function, EI(x) is the function to be improved, and x * is the location of the optimal sampling point, f(x * ) is the current optimal solution.

9. The multi-feature power load forecasting method based on VMD-BO-LSTM according to claim 1, characterized in that: The step S60 comprises: Input the core feature factor data of the test set into the trained LSTM network model; Reconstruct the K intrinsic mode function values ​​output by the long short-term memory network model to obtain the predicted load; The indicator evaluation is carried out based on the predicted load and the daily load data of the test set.

10. The multi-feature power load forecasting method based on VMD-BO-LSTM according to claim 9, characterized in that: The specific indicators of the indicator evaluation include: MAE, RMSE, MAPE, R 2 and NSE.

Images