Power load prediction method based on DWT-CNN-LSTM model

Abstract

The invention discloses a power load prediction method based on a DWT-CNN-LSTM model, and the method comprises the steps: carrying out the decomposition of collected original load data through employing a discrete wavelet decomposition method, obtaining a low-frequency load component and a high-frequency load component, carrying out the denoising of the high-frequency component on the basis, and not carrying out the special processing of the low-frequency component; inputting the load data of different components into a convolutional neural network, and extracting features of the load data of different components; and respectively substituting each load component after convolution into the LSTM model for training to obtain an LSTM load prediction model of each load component, and then carrying out sequence reconstruction on the load prediction value of each component to obtain the prediction value of the power load. According to the method, the problem that noise exists in signals can be effectively solved, and power load data can be accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system load artificial intelligence technology prediction, and specifically relates to a power load prediction method based on a DWT-CNN-LSTM model. Background Art

[0002] Accurate power load forecasting is conducive to the safe and economic operation of the power grid, provides basic data basis for power grid planning, plays an important role in the development of modern power systems, and has a decisive impact on the quality of power grid planning. How to improve the accuracy of load forecasting has always been a hot topic for scholars.

[0003] In recent years, artificial intelligence technology has been widely used in power load forecasting scenarios due to its advantages in nonlinear fitting. With the development of smart grids, grid data has become richer, and more and more influencing factors have been considered. However, the data of influencing factors have more and more time scales, finer data granularity, and non-equal time intervals. It is necessary to explore the processing methods of multi-time scale data to integrate these data into a model.

[0004] The development trend of power load forecasting methods is to take more and more factors into consideration. In the case of multiple factors interacting with each other, the method develops from analytical methods to non-analytical and intelligent methods. The current commonly used load forecasting methods have the following problems: With the rapid development of smart grids, the composition of power grids has become richer, and the factors affecting power loads also appear on more time scales, and their influence on loads is inconsistent. The deeper the neural network, the stronger its expression ability, allowing to find approximate mappings of real laws in a larger function space. With the assistance of a suitable optimization method, the neural network can converge to a good prediction effect. Summary of the invention

[0005] In order to solve the above technical problems and consider the importance of factors at different time scales, the present invention provides a method for predicting power load based on the DWT-CNN-LSTM model. First, the historical load data is decomposed into different frequency scales using discrete wavelet decomposition. In order to enable the prediction model to better grasp the influence of long-, medium- and short-term factors on the power load, a convolutional neural network is introduced into the model to suppress noise interference while extracting the characteristics of load data of different components. Then, the long short-term memory neural network model is used for training to further improve the prediction accuracy of the model. This method can accurately predict power load data under the condition of multiple factors interacting with each other. The present invention can provide important data for the power department to formulate maintenance plans and make decision-making plans, which is of great significance for ensuring the safe and economical operation of the power system.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: A method for power load forecasting based on a DWT-CNN-LSTM model includes the following steps: Step S1: Obtain the original power load data, obtain the original load sequence, use discrete wavelet transform to decompose the original load sequence into a low-frequency sequence and a high-frequency sequence, and perform noise reduction on the high-frequency component D of the high-frequency sequence; Step S2: input the low-frequency component A and the high-frequency component D after discrete wavelet decomposition and denoising into CNN respectively to learn and extract the features of each component; Step S3: After being processed by CNN, each component data enters the LSTM model for training; Step S4: The load values ​​predicted by the LSTM model for each component are sequenced and the reconstructed result is the final power load prediction value.

[0007] Beneficial effects: 1) The present invention uses discrete wavelet transform to decompose the original power load data into low-frequency components and high-frequency components, and introduces a convolutional neural network to learn and extract the characteristics of each load component, so that the model can pay attention to the characteristics of different scales during training; 2) The present invention adopts different neural network models to fit and predict the load time series components at different scales, processes the nonlinear part of the wavelet coefficients with a convolutional neural network (CNN), and processes the volatility part with an LSTM neural network, so that the prediction accuracy is significantly improved; 3) Normalize the sample data so that the neural network can converge quickly and avoid neuron output saturation; 4) The method of the present invention can accurately predict power load data under the influence of multiple factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flow chart of a method for power load forecasting based on a DWT-CNN-LSTM model of the present invention.

[0009] Figure 2 This is the flow chart of load data preprocessing.

[0010] Figure 3 This is the load forecasting flow chart.

[0011] Figure 4 Schematic diagram of wavelet decomposition. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0013] like Figure 1 As shown, the power load forecasting method based on the DWT-CNN-LSTM (discrete wavelet transform-convolutional neural network-long short-term memory network) model of the present invention includes the following steps: Step S1: Obtain the original power load data, obtain the original load sequence, use discrete wavelet transform to decompose the original load sequence into a low-frequency sequence and a high-frequency sequence, and perform noise reduction on the high-frequency component D of the high-frequency sequence; Step S2: inputting the low-frequency component A and the high-frequency component D after discrete wavelet decomposition and denoising into CNN respectively to learn and extract the features of each component; Step S3: After being processed by CNN, each component data enters the LSTM model for training; Step S4: The load values ​​predicted by the LSTM model for each component are sequenced and the reconstructed result is the final power load prediction value.

[0014] Specifically, Figure 2 As shown, the step S1 includes: Power load data is the result of the combined effects of multiple complex factors, and a single prediction model cannot directly reflect the nonlinear characteristics of the corresponding sequence. Wavelet transform achieves multi-scale decomposition of the signal by scaling and translating it, which can display the frequency information of each component of the target original signal and is very suitable for analyzing non-stationary signals.

[0015] To this end, the present invention uses wavelet transform to decompose the original number into different frequency scales before load forecasting, and normalizes the data to complete the preprocessing of the load data.

[0016] Wavelet decomposition can effectively separate the components with characteristic differences in the signal while retaining the time domain information, that is, each layer of wavelet decomposition will obtain high-frequency components. and low frequency components Two parts, for The low-frequency component after decomposition is for The high-frequency components after decomposition, such as Figure 4 shown.

[0017] Discrete wavelet transform is to decompose the original time series X into signals of different frequencies through a series of half-bandwidth low-pass filters H and high-pass filters G: ; In the formula, is the current wavelet decomposition level, is the total number of wavelet decomposition layers; is the time sequence number, i.e. wavelet coefficients; is the low-pass filter coefficient, is the high-pass filter coefficient, for The dimensional signal is Approximate components decomposed by layers; for The dimensional signal is The detail components decomposed into layers. for The approximate components of the decomposed dimensional signal at the first layer; for The dimensional signal is Low-pass filter coefficients on dimensional time series; The original signal is The size of the dimensional time series; for The approximate components of the decomposed dimensional signal at the first layer; for The dimensional signal is High-pass filter coefficients on a time series; For the dimensional time series in Approximate components decomposed by layers; The output after wavelet decomposition is: ; In the formula, is the original load data; For the The first-dimensional approximate component decomposed from the layer; For the The second-dimensional approximate component decomposed from the layer; For the The decomposition of the layers dimensional approximate component; It is the first-dimension detail component decomposed from the first layer; It is the second-dimensional detail component decomposed from the first layer; is the decomposition of the first layer Dimensional detail component; For the The first dimension detail component decomposed from the layer; For the The second-dimensional detail component decomposed from the layer; For the The decomposition of the layers Dimensional detail.

[0018] In step S1, the relationship between the components of each layer is as follows: ; ; ; · · · ; In the formula, Represents the original data, Indicates The low-frequency component coefficients of layer decomposition, Indicates High frequency component coefficients of layer decomposition; =1, 2, 3···n.

[0019] Preferably, step S1 further includes normalizing the historical load data, and the calculation formula is: ; In the formula, Indicates historical load data; , Respectively represent the minimum and maximum values ​​of historical load data; Represents normalized historical load data.

[0020] Specifically, Figure 3 As shown, step S2 includes: The convolutional neural network (CNN) in step S2 is a type of feedforward neural network with deep structure and convolution calculation, and is one of the representative algorithms of deep learning. CNN can dynamically learn features from raw data without manually designing feature extractors, which is particularly beneficial for high-dimensional and complex time series data such as power load data. Power load data often has complex spatiotemporal relationships and dependencies. CNN can abstract the features of data layer by layer through multi-layer convolution and pooling operations, thereby capturing different levels and patterns in the data, which makes CNN more effective than traditional methods in capturing the intrinsic structure of data.

[0021] Preferably, in step S2, the CNN neural network includes a convolution layer, an activation function, a pooling layer and a fully connected layer, wherein the convolution layer uses a set of learnable filters to scan the input data, and each filter is responsible for extracting a specific feature; the activation function introduces a Sigmoid nonlinear function to enable the network to learn complex patterns; the pooling layer reduces the spatial dimension of the feature map, reduces the amount of calculation and the number of parameters, and prevents overfitting; the fully connected layer converts the feature map extracted by the previous convolution layer and the pooling layer into a feature vector so that the feature value can be input into the subsequent LSTM model.

[0022] 1. The convolutional layer is: ; The summation part in the formula is equivalent to solving a cross-correlation, is the deviation, and Indicates The convolutional input and output of the layer, Corresponding to Pixels of the layer feature map; , are the length and width of the feature vector, For the The convolution weight matrix of the layer, Indicates Dimensional signal, For the layer The convolution weight matrix of the dimensional signal, Used for convolution kernel length counting, Used for convolution kernel width counting, For the Layer The feature map of the dimensional signal, For the The number of channels of the layer, , are the convolution layer parameters, corresponding to the convolution kernel size and convolution step size; 2. Pooling layer: The pooling layer is another important component in CNN, which is responsible for reducing the dimension of the feature map, reducing the amount of calculation and the risk of overfitting. The pooling operation usually adopts the maximum pooling or average pooling method to divide the feature map into multiple sub-regions, and then select the maximum value or average value of each sub-region as the output. This can effectively reduce the dimension of the feature map while retaining important feature information. The formula of the pooling layer is as follows: ; In the formula, and Indicates layer dimensional signal pooling output and input, convolution kernel size , Step Length , Pixels , convolution kernel length count variable and convolution kernel width count variables The meaning is the same as that of the convolutional layer. is a pre-specified parameter. hour, Pooling takes the average value in the pooling area, which is called average pooling. hour, Pooling takes the maximum value in the region, which is called max pooling. Mean pooling and max pooling are pooling methods that have been used for a long time in the design of convolutional neural networks. Both retain the background and texture information of the image at the cost of losing some information or size of the feature map.

[0023] 3. Activation function: The result of the convolution operation is a short array, in which each value is obtained by multiplying the input data with the corresponding element of the convolution kernel and then summing it. In order to enable the network to learn nonlinear features, activation functions are introduced in CNN. Common activation functions include RelU, Siqmoid and tanh. These functions can convert input values ​​into nonlinear forms, so that the network can better learn and recognize complex features. The present invention uses Sigmoid function as the activation function, and its formula is as follows: ; in, is the output of the Sigmoid function, Represents the input of the activation function. The Sigmoid function is a common S-shaped function in biology, also known as the S-shaped growth curve. In deep learning, due to its monotonic and inverse monotonic properties, the Sigmoid function is often used as an activation function of a neural network to map variables to between [0, 1].

[0024] 4. Fully connected layer: At the final stage of a CNN, there is usually one or more fully connected layers, also called dense layers or output layers. These layers are responsible for combining previously extracted features for tasks such as classification or regression. Each neuron in a fully connected layer is connected to all neurons in the previous layer and uses common activation functions such as Softmax for classification or regression prediction. The formula for a fully connected layer is as follows: ; In the formula, is the output vector of the fully connected layer, is the input vector of the fully connected layer, is the activation function, is the weight matrix, is the bias vector.

[0025] The LSTM neural network includes a forget gate, an input gate, an output gate and a memory unit neuron. The forget gate takes the input at the current moment and the output at the previous moment as input, and controls the extent to which the unit state at the previous moment is transmitted to the current moment; the input gate controls how much new information is added, and works with the forget gate to update the unit state of the memory unit; the output gate controls the extent to which the current unit state is deleted.

[0026] LSTM neural network is a special case of recurrent neural network, which can avoid the vanishing gradient of conventional RNN and learn time series feature information well. The high volatility coefficient decomposed from load series data is learned through LSTM neural network model to predict its unconventional fluctuations and obtain more accurate results.

[0027] Specifically, step S3 includes: Step S3.1. Training the LSTM model includes: The preprocessed data is divided into two groups, T1 and T2. Group T1 is used as training samples, and group T2 is used as test samples. The number of training sample groups T1 is generally larger than the number of test sample groups T2. For example, the ratio can be 70% and 30%. The training samples are used as the input of the LSTM network, and the test samples are used to test the performance of the network model.

[0028] Step S3.1. Training the LSTM model includes: The preprocessed data is divided into two groups, T1 and T2. Group T1 is used as training samples, and group T2 is used as test samples. The number of training sample groups T1 is generally larger than the number of test sample groups T2. For example, the ratio can be 70% and 30%. The training samples are used as the input of the LSTM network, and the test samples are used to test the performance of the network model.

[0029] Step S3.1.1: Network hyperparameter initialization: The hyperparameters that need to be set manually in the present invention include: the number of input nodes , the number of hidden nodes , output node number , error threshold , maximum number of iterations , the number of neuron nodes in the two hidden layers of LSTM , , and the learning rate .

[0030] Step S3.1.2: Weight bias initialization: The forward calculation formula of the LSTM neural network is as follows: ; The above formula indicates The forward calculation formula of LSTM at this moment can obtain the output value of each neuron. , , , , Respectively The values ​​of the five vectors at the moment are the forget gate output, input gate output, output gate output, unit state and load output value. Through forward calculation, LSTM can more effectively determine which information is forgotten and which information is retained, and process long time series more efficiently. , , for The three inputs of the network at this moment are: is the final output of the network. for The load output value at the moment is dimensional row vector, Indicates the number of loads that need to be predicted, that is, the number of output nodes; for The network input value at time , which is a dimensional row vector, is the dimension of a single sample, that is, the number of input nodes; is the current input unit status; for The unit state of the network at that moment; for The electrical load output value at the moment is also a dimensional row vector. , , , Respectively The forget gate output, input gate output, output gate output and unit state at the moment. and Are all activation functions, among which, is the sigmoid function, is the hyperbolic tangent function, and the calculation formulas are as follows: ; ; In the formula, is the input vector.

[0031] “ " means element-wise multiplication. When " "When applied to two vectors, the operation is as follows: ; when" "When applied to a vector and a matrix, the operation is as follows: ; when" When applied to two matrices, the elements in corresponding positions of the two matrices are multiplied.

[0032] in, , , , Represent the weight matrices of the forget gate, input gate, current input unit state, and output gate, respectively. , , , It represents the bias matrix of the forget gate, input gate, current input unit state and output gate. These 8 parameters are trained by the network and do not need to be set manually. However, the matrix dimension needs to be specified manually and a random number between 0 and 1 is generated by the computer as its initial value.

[0033] Step S3.1.3: According to the loss function Update network weights and biases; loss function The calculation formula is as follows: ; In the formula, represents the total number of training samples, Indicates The network output voltage or internal resistance value of the training sample, Indicates The actual voltage or internal resistance value corresponding to each training sample. By changing the weight bias value, The size of the gradient descent algorithm can quickly find the appropriate weight and bias, so that Minimize to reach the error threshold , speeding up network convergence.

[0034] Step S3.1.4: Repeat step S3.1.3 until the number of iterations equals Or the error is less than the error threshold , stop training.

[0035] Step S3.1.5: Add the predicted values ​​of each component of DWT to obtain the final network load prediction result.

[0036] Step S3.2. LSTM accuracy verification: After the LSTM network training is completed, the test sample T2 is used for testing, and the final predicted value of the network model is compared with the true value. The present invention uses statistical error to evaluate the performance of the LSTM neural network. The statistical error used in this method is the root mean square error , the calculation formula is as follows: ; In the formula, For the data set The network prediction value for observations, For the data set The actual value of the observations, is the number of network prediction values. If the statistical error meets the actual needs, the LSTM model is accurate and can be put into practice. If it does not meet the actual needs, the network model needs to be retrained.

[0037] Specifically, step S4 includes: based on the decomposed wavelet coefficients of each layer , , , , , , , the predicted values ​​of the low-frequency sequence and the high-frequency sequence are integrated to reconstruct the predicted value of the original load sequence.

[0038] In summary, the present invention proposes a power load forecasting method based on the DWT-CNN-LSTM model for complex power load data. Through discrete wavelet decomposition and normalization processing, the interference of high-frequency and low-amplitude noise in the training samples is reduced, and the focus on important components is improved. The load data of different components are input into CNN to extract the characteristics of the load data of different components; each load component after convolution is substituted into the LSTM model for training to obtain the LSTM load forecasting model of each load component. The method proposed in the present invention has good adaptability and accuracy for medium- and long-term load forecasting, which is an important basis for the power department to formulate maintenance plans and make decision-making plans, and is of great significance for ensuring the safe and economical operation of the power system.

Claims

1. A power load forecasting method based on a DWT-CNN-LSTM model, comprising the following steps: Step S1: Obtain the original power load data, obtain the original load sequence, use discrete wavelet transform to decompose the original load sequence into a low-frequency sequence and a high-frequency sequence, and perform noise reduction on the high-frequency component D of the high-frequency sequence; Step S2: input the low-frequency component A and the high-frequency component D after discrete wavelet decomposition and denoising into CNN respectively to learn and extract the features of each component; Step S3: After being processed by CNN, each component data enters the LSTM model for training; Step S4: The load values ​​predicted by the LSTM model for each component are sequenced and the reconstructed result is the final power load prediction value.

2. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 1 is characterized in that: In step S1, decomposing the original load sequence by discrete wavelet transform includes the following steps: Step S11: decompose the original load sequence into high-frequency components; Step S12: perform binary downsampling on the low-frequency component to obtain DWT coefficients, continue to perform decomposition iterations and gradually approximate, and finally obtain a low-frequency component and multiple high-frequency components; DWT stands for discrete wavelet transform.

3. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 2 is characterized in that: In step S1, the relationship between the components of each layer is as follows: ； ； ； · · · ； In the formula, Represents the original data, Indicates The low-frequency component coefficients of layer decomposition, Indicates High frequency component coefficients of layer decomposition; =1, 2, 3···n.

4. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 1 is characterized in that: In step S2, the CNN neural network includes a convolution layer, an activation function, a pooling layer and a fully connected layer. The convolution layer uses a group of learnable filters to scan the input data, and each filter is responsible for extracting a specific feature; the activation function introduces the Sigmoid nonlinear function to enable the network to learn complex patterns; the pooling layer reduces the spatial dimension of the feature map, reduces the amount of calculation and the number of parameters, and prevents overfitting; the fully connected layer converts the feature map extracted by the previous convolution layer and the pooling layer into a feature vector so that the feature value can be input into the subsequent LSTM model.

5. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 2 is characterized in that: In step S2, the mathematical expression of the CNN neural network is: The convolutional layer is: ； The summation part in the formula is equivalent to solving a cross-correlation, is the deviation, and Indicates The convolutional input and output of the layer, Corresponding to Pixels of the layer feature map; , are the length and width of the feature vector, For the The convolution weight matrix of the layer, Indicates Dimensional signal, For the layer The convolution weight matrix of the dimensional signal, Used for convolution kernel length counting, Used for convolution kernel width counting, For the Layer The feature map of the dimensional signal, For the The number of channels of the layer, , are the convolution layer parameters, corresponding to the convolution kernel size and convolution step size; Pooling layer: The pooling layer is another important component in CNN, which is responsible for reducing the dimension of the feature map, reducing the amount of calculation and the risk of overfitting; the pooling operation usually adopts the maximum pooling or average pooling method, divides the feature map into multiple sub-regions, and then selects the maximum value or average value of each sub-region as the output, which can effectively reduce the dimension of the feature map while retaining important feature information; the formula of the pooling layer is as follows: ； In the formula, and Indicates layer dimensional signal pooling output and input, convolution kernel size , Step Length , Pixels , convolution kernel length count variable and convolution kernel width count variables The meaning is the same as that of the convolutional layer. is a pre-specified parameter; when hour, Pooling takes the average value in the pooling area, which is called average pooling. hour, Pooling takes the maximum value in the region, which is called max pooling. Mean pooling and max pooling are pooling methods that have been used for a long time in the design of convolutional neural networks. Both retain the background and texture information of the image at the cost of losing some information or size of the feature map. Activation function: The result of the convolution operation is a short array, in which each value is obtained by multiplying the input data with the corresponding element of the convolution kernel and then summing them. In order to enable the network to learn nonlinear features, an activation function is introduced in CNN. Common activation functions include RelU, Siqmoid and tanh, etc. These functions can convert the input value into a nonlinear form, so that the network can better learn and recognize complex features; the present invention adopts the Sigmoid function as the activation function, and its formula is as follows: ； in, is the output of the Sigmoid function, Represents the input of the activation function. The Sigmoid function is a common S-shaped function in biology, also known as the S-shaped growth curve. In deep learning, due to its monotonic and inverse monotonic properties, the Sigmoid function is often used as the activation function of a neural network to map variables to between [0, 1]. Fully connected layer: In the final stage of CNN, there is usually one or more fully connected layers, also called dense layers or output layers; these layers are responsible for combining the previously extracted features for tasks such as classification or regression. Each neuron in the fully connected layer is connected to all neurons in the previous layer and uses common activation functions such as Softmax for classification or regression prediction; the formula for the fully connected layer is as follows: ； In the formula, is the output vector of the fully connected layer, is the input vector of the fully connected layer, is the activation function, is the weight matrix, is the bias vector.

6. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 1 is characterized in that: In step S3, the LSTM model includes a forget gate, an input gate, an output gate and a memory unit neuron. The forget gate takes the input at the current moment and the output at the previous moment as input, and controls the extent to which the unit state at the previous moment is transmitted to the current moment; the input gate controls the amount of new information added, and cooperates with the forget gate to update the unit state of the memory unit; the output gate controls the extent to which the current unit state is deleted.

7. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 6 is characterized in that: The step S3 comprises: Step S3.

1. Training the LSTM model includes: The data obtained after preprocessing are divided into two groups, T1 and T2. Group T1 is used as training samples, and group T2 is used as test samples. The number of training sample groups T1 is greater than the number of test sample groups T2. Step S3.1.1: Network hyperparameter initialization: The artificially set hyperparameters include: number of input nodes , the number of hidden nodes , output node number , error threshold , maximum number of iterations , the number of neuron nodes in the two hidden layers of LSTM , , and the learning rate ; Step S3.1.2: Weight bias initialization: The forward calculation formula of the LSTM neural network is as follows: ； The above formula indicates The forward calculation formula of LSTM at this moment, the forward calculation can get the output value of each neuron, , , , , Respectively The values ​​of the five vectors at the moment: forget gate output, input gate output, output gate output, unit state and electrical load output value; , , for The three inputs of the network at this moment are: is the final output of the network, where for The load output value at the moment is dimensional row vector, Indicates the number of loads that need to be predicted, that is, the number of output nodes; for The network input value at time , which is a dimensional row vector, is the dimension of a single sample, that is, the number of input nodes; is the current input unit status; for The unit state of the network at that moment; for The electrical load output value at the moment is also a dimensional row vector; , , , Respectively The forget gate output, input gate output, output gate output and unit state at the moment; and Are all activation functions, among which, is the sigmoid function, is the hyperbolic tangent function, and the calculation formulas are as follows: ； ； In the formula, is the input vector; " means element-wise multiplication, when " "When acting on two matrices, the elements of the two matrices at corresponding positions are multiplied; in, , , , Represent the weight matrices of the forget gate, input gate, current input unit state, and output gate, respectively. , , , It represents the bias matrix of the forget gate, input gate, current input unit state and output gate. These 8 parameters are trained by the network and do not need to be set manually. However, the matrix dimensions need to be specified manually and a random number between 0 and 1 is generated by the computer as its initial value. Step S3.1.3: According to the loss function Update network weights and biases; loss function The calculation formula is as follows: ； In the formula, represents the total number of training samples, Indicates The network output voltage or internal resistance value of the training sample, Indicates The actual voltage or internal resistance value corresponding to the training samples; Step S3.1.4: Repeat step S3.1.3 until the number of iterations equals Or the error is less than the error threshold , stop training; Step S3.1.5: Add the predicted values ​​of each component of DWT to obtain the final network load prediction result; Step S3.

2. LSTM accuracy verification: After the LSTM network training is completed, the T2 group is used as the test sample for testing. The final predicted value of the network model is compared with the true value, and the root mean square error is used. To evaluate the performance of the LSTM neural network, the calculation formula is as follows: ； In the formula, For the data set The network prediction value for observations, For the data set The actual value of the observations, is the number of network prediction values.

8. A method for predicting power load based on a DWT-CNN-LSTM model according to any one of claims 1 to 7, characterized in that: Step S1 also includes normalizing the historical load data, and the calculation formula is: ； In the formula, Indicates historical load data; , Respectively represent the minimum and maximum values ​​of historical load data; Represents normalized historical load data.

9. The power load forecasting method based on the DWT-CNN-LSTM model according to claim 5 is characterized in that: In step S4, based on the decomposed wavelet coefficients of each layer , , , , , the predicted values ​​of the low-frequency sequence and the high-frequency sequence are integrated to reconstruct the predicted value of the original load sequence.

Images