CNN-LSTM-based high-voltage user power consumption data prediction method

Abstract

The invention discloses a CNN-LSTM-based high-voltage user power consumption data prediction method, which takes high-voltage user power consumption data as a research object, cleans historical power consumption data of high-voltage users by using an LOF method, removes unreasonable outlier data, and reduces the influence of the outlier data on the performance of a prediction model. Through data classification and feature engineering, original data are converted into a more representative feature set, which is convenient for model learning. And the volatility of the data is reduced through data normalization, so that the stability and convergence speed of model training are improved. And then constructing a CNN-LSTM hybrid model, that is, adding a convolutional neural network CNN on the basis of a long short-term memory network LSTM. The hybrid model is composed of a CNN part and an LSTM part, the CNN part is used for extracting local features of data, the LSTM part is used for capturing a dependency relationship of a time sequence, and then more accurate power consumption prediction is carried out. After the construction and training of the model are completed, the prediction result of the model is effectively and intuitively evaluated through a mean square error and a graphical method.

Description

Technical Field

[0001] The present invention relates to the field of power grid data prediction, and specifically proposes a method for predicting high-voltage user electricity consumption data based on CNN-LSTM. Background Art

[0002] High-voltage users are usually the main components of large-scale industrial, manufacturing and commercial sectors. The production and operation activities of these users are crucial to the country's economic growth and competitiveness. Stable power supply ensures the production capacity of these key sectors, helps maintain national economic stability and improve the country's overall competitiveness. Stable power supply is inseparable from scientific and accurate power data forecasting.

[0003] At present, according to the different periods of power data forecast, power data forecast can be divided into long-term forecast, medium-term forecast, short-term forecast and ultra-short-term forecast, corresponding to annual forecast, monthly forecast, daily forecast and hourly forecast respectively. Forecasts in some areas also include ten-day forecast and weekly forecast, but they are not very common. According to different forecast mechanisms, annual forecast and monthly forecast are collectively referred to as medium- and long-term forecast, and daily forecast and hourly forecast are classified as short-term forecast. In addition, in each period, the content of power consumption forecast generally includes load (load curve) and power consumption forecast.

[0004] According to the prediction methods of power data, the research on power data prediction can be divided into two categories: one is the traditional mathematical prediction method, including regression analysis, grayscale prediction, exponential smoothing method and differential autoregressive moving average method; the other is the prediction method based on artificial intelligence, such as tree ensemble algorithm, support vector machine, neural network algorithm, etc. The first type of method is also called traditional method, which usually requires domain experts to select and design features, which may require a lot of time and expertise, and traditional methods are usually based on linear assumptions. Due to the complexity of practical problems, it is sometimes difficult or even impossible to capture such linear relationships. The prediction methods based on artificial intelligence, especially deep learning, can automatically learn complex patterns and associations from large-scale data. This allows such prediction algorithms to better cope with nonlinear, multivariate and elusive data relationships. In addition, deep learning can automatically learn to extract useful features from data without manual feature engineering.

[0005] LSTM neural network is a specific form of recurrent neural network (RNN). The LSTM model solves the problem of RNN short-term memory by adding gates on the basis of the RNN model, so that the recurrent neural network can truly and effectively utilize long-distance time series information. In addition, LSTM neural networks are widely used in time series prediction. In the prediction of electricity consumption, many scholars have tried to use the LSTM neural network model. And after many experiments, it has been shown that the LSTM neural network model has a high reliability in the prediction accuracy of the electricity consumption model. Hamzehei et al. predicted electricity consumption by using LSTM neural network. The study is mainly based on the current electricity demand, including some influencing factors such as temperature, humidity and seasonality. By analyzing the prediction results, it is found that the LSTM model has good prediction accuracy and robustness. Summary of the invention

[0006] The purpose of the present invention is to provide a method for predicting electricity consumption data of high-voltage users based on CNN-LSTM. It is necessary to summarize the historical electricity consumption data of high-voltage users in a certain area and pre-process the original data. Remove unreasonable data, perform feature engineering, convert the data into a format suitable for LSTM model input, and normalize the data to reduce the volatility of the data. Construct a CNN-LSTM hybrid model, model the pre-processed data using the established hybrid model algorithm, and select a good model evaluation method to evaluate the model.

[0007] A method for predicting high-voltage user electricity consumption data based on CNN-LSTM, comprising the following steps:

[0008] S1: Data processing of historical electricity consumption data of high-voltage users, including data cleaning, data classification and feature engineering, data normalization, data splitting and sequence creation;

[0009] S2: Use the processed high-voltage user data to build a CNN-LSTM hybrid model; add a convolution layer before the LSTM network to extract local features in the data, and then build a two-layer LSTM neural network model, each layer contains 64 LSTM units; and use the training set data to train the model;

[0010] S3: Use error metrics and graphical methods to evaluate the prediction effect of the model; put the evaluated model into use and input data to obtain the corresponding prediction results.

[0011] The beneficial effects achieved by the present invention are:

[0012] (1) Strong long-term dependency capability: The electricity consumption data of high-voltage users, the research object of the present invention, has a long time span. Conventional neural network models are prone to gradient vanishing problems in the process of processing such long sequence data, resulting in reduced accuracy and reliability of the model. The present invention adopts the LSTM prediction method, which makes full use of its unique memory unit structure and the ability to effectively capture and retain long-term dependent information, allowing the network to remember or forget information when necessary, preventing gradient vanishing and gradient explosion problems, thereby ensuring good prediction accuracy when predicting electricity consumption data of long-time span sequences.

[0013] (2) Rapid model generation: Traditional mathematical methods are usually based on linear assumptions. Due to the complexity of high-voltage user electricity consumption data, it is sometimes difficult or even impossible to capture this linear relationship. This type of method usually requires domain experts to select and design features, which may require a lot of time and expertise. Compared with conventional neural network models, LSTM has good generalization ability and can quickly train new models on new data sets. This is due to the fact that LSTM retains important feature information through its structural memory unit and can quickly adjust its weight parameters to adapt to new data distribution when faced with new data. Specifically, LSTM's ability to quickly train on new data sets is reflected in its gradient update algorithm and gating mechanism, which enables it to adjust parameters in a short time and converge to a new optimal solution, thereby achieving fast and accurate predictions. In addition, the present invention reduces the volatility of the data through data normalization operations, so that different features are on the same scale, thereby improving the stability and convergence speed of model training.

[0014] (3) Good prediction accuracy: The LSTM neural network model can accurately predict future electricity consumption based on historical data and related environmental factors. The results show that compared with traditional time series analysis methods, the LSTM neural network model can significantly improve the prediction accuracy. In terms of monthly and weekly time span predictions, the error between the predicted electricity consumption and the actual value is less than 2%. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the steps of data cleaning in an embodiment of the present invention.

[0016] Figure 2 It is a flow chart of constructing a CNN-LSTM hybrid model in an embodiment of the present invention.

[0017] Figure 3 It is a flowchart of the steps of model training in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0019] The method of the present invention is to aggregate the historical power consumption data of high-voltage users, use the LOF (Local Outlier Factor, local outlier factor) method to clean the original data and remove unreasonable outlier data. Then the data is classified and accumulated for the region, plate, industry, group, etc. to which the high-voltage users belong, and the data is decomposed into information of different time granularities according to the month, week, and day according to the date feature, and other possible feature engineering processes are performed. The features are normalized to reduce the volatility of the data. The data is split into a training set and a test set, and converted into a time series format suitable for LSTM model input. A CNN-LSTM hybrid model is constructed, that is, one or more convolutional layers are added before the LSTM network to extract local features in the data, and then the time series features are captured by the LSTM network. This architecture can combine time series features and spatial features to more effectively predict power consumption data. The preprocessed data is modeled using the established LSTM algorithm, and the time series prediction data of the corresponding time granularity in the future is input. The mean square error (MSE) and graphical methods (comparing the prediction result curve with the time change and the measured results) are used to evaluate the prediction accuracy of the model.

[0020] Electricity consumption forecasting is of great significance in energy management and planning. Accurate forecasting helps optimize the allocation of power resources and ensure the stable operation of the power system. This paper combines LSTM neural network and data preprocessing technology to propose an efficient electricity consumption forecasting method.

[0021] This paper proposes a method for predicting power consumption of high-voltage users based on long short-term memory (LSTM) neural network. The historical power consumption data is cleaned, classified and feature-engineered, and the LOF (local outlier factor) method is used to remove unreasonable outlier data, and finally the data is input into the LSTM model for prediction. The accuracy of this method is evaluated by mean square error (MSE) and graphical method.

[0022] The steps of the prediction method are as follows.

[0023] Step 1: Data processing.

[0024] 1.1) Data cleaning. The LOF method is used to clean the historical power consumption data of high-voltage users and remove unreasonable outlier data. The LOF algorithm identifies outliers by comparing the density difference between data points and their neighbors, thereby ensuring the validity and accuracy of the data.

[0025] 1.2) Data classification and feature engineering. Classify and accumulate data according to the region, sector, industry, group, etc. to which the high-voltage users belong. Further decompose the date features into information of different time granularities such as month, week, and day, and perform feature engineering processing, including adding features such as year, month, day, and day of the week.

[0026] 1.3) Data normalization. Normalize the features to reduce the volatility of the data and ensure the stability and convergence of the model training.

[0027] 1.4) Data Splitting and Sequence Creation: Split the data into training and test sets, and convert them into a time series format suitable for LSTM model input.

[0028] Step 2: Model building and training

[0029] 2.1) Construct a CNN-LSTM hybrid model. Add one or more convolutional layers before the LSTM network to extract local features in the data, and then construct a two-layer LSTM neural network model, each layer contains 64 LSTM units, followed by a hidden layer using the ReLU activation function, and finally an output layer using a linear activation function. The mean square error MSE is used as the loss function. The Adam optimizer is used for model training; the Adam optimizer combines the advantages of momentum and RMSProp and accelerates gradient descent by adaptively adjusting the learning rate.

[0030] 2.2) Model training. Use the training set data to train the model and set the appropriate number of training rounds and batch size.

[0031] Step 3: Model Evaluation

[0032] 3.1) Error metric: The mean square error (MSE) is used to evaluate the prediction results of the model.

[0033] 3.2) Graphical method: By drawing the time series graph and residual graph of the predicted value and the true value, the accuracy of the model prediction results can be intuitively evaluated.

[0034] The proposed method for predicting power consumption of high-voltage users effectively improves data quality through data cleaning, feature engineering and normalization, and uses the LSTM model to accurately predict power consumption. The model is evaluated by mean square error and graphical methods to verify the accuracy and stability of the model. The specific construction information is shown in the method flow.

[0035] For some steps in the above prediction method flow, the specific implementation methods are as follows.

[0036] The purpose of data cleaning is to remove outliers in the data set to ensure the quality and consistency of the data. The Local Outlier Factor (LOF) algorithm is used to detect and remove outliers.

[0037] The LOF algorithm detects outliers through the following steps:

[0038] Distance calculation:

[0039] For each data point x in the dataset i , calculate its difference with all other data points x j The Euclidean distance of:

[0040] d(x i , x j )=||x i -x j ||

[0041] k-distance neighborhood:

[0042] For a data point x i , find its k nearest neighbor data points, called k-neighborhood, denoted by N k (x i ).

[0043] Reachable distance:

[0044] Define data point x j For data point x i The reachable distance is:

[0045] reach-dist k (x i , x j )=max{d(x i , x j ), k-distance(x j )}

[0046] Among them, k-distance(x j ) is the kth nearest neighbor to x j The local reachability density is:

[0047] Define data point x i The local reachable density is:

[0048]

[0049] Among them, N k (x i ) is the data point x i k-neighborhood

[0050] Local outlier factor:

[0051] Finally, define the data point x i The local outlier factor of is:

[0052]

[0053] When LOF k (x i ) value is much greater than 1, the data point x i are considered outliers.

[0054] The purpose of data classification and feature engineering is to convert raw data into a more representative feature set to facilitate model learning.

[0055] The data is classified according to the region, sector, industry, group, etc. to which the high-voltage users belong. Assume that the data set D contains n data points, each of which has m features. According to the classification features, the data set is divided into C subsets {D1, D2, ..., D C}.

[0056] Decompose the date feature into month, week, and day information. Assuming the date feature is t, the following features are generated:

[0057] month m = month(t);

[0058] Week w = dayofweek(t);

[0059] ·Day d=day(t).

[0060] The original data and the generated date features are further combined to enhance the feature expression ability of the model and complete data classification.

[0061] Combined feature f new =f i ×f j ;

[0062] Aggregate feature f agg =∑ k f k .

[0063] Among them, the combined feature is a new feature generated by performing some combination operation on the existing features. For example, if we have two original features f i and f j , a new combined feature f can be generated through multiplication, addition or other operations new ; Aggregate feature f agg is obtained by computing a set of features {f1, f2, ..., f k} Features obtained by performing aggregation operations (such as sum, average, maximum, etc.).

[0064] The purpose of data normalization is to reduce the volatility of data and make different features on the same scale, thereby improving the stability and convergence speed of model training. The commonly used method is Z-score normalization.

[0065] For each feature v i After standardization, the formula is:

[0066]

[0067] Among them, μ i and σ i They are the features v i The mean and standard deviation of .

[0068] The data is split into training and testing sets and converted into a time series format suitable for LSTM model input.

[0069] The standardized data set is divided into the training set D according to the proportion p train and the test set D test Here p is usually set to 4, that is, 80% of the data is used as the training set and 20% as the test set.

[0070] Assume that the time step is T and convert the data set into a time series format. Each time series sample contains T time steps of data. Define the time series sample S i for:

[0071] S i = {V i , V i+1 , ..., V i+T-1}

[0072] Among them, V i is the eigenvector at time step i.

[0073] The process of building the CNN-LSTM hybrid model includes the design of 5.1 convolutional layer, LSTM unit structure, two-layer LSTM network, loss function, and optimizer.

[0074] In order to capture the local features of high-voltage user electricity consumption data, a convolutional layer is added before LSTM. The specific convolutional layer design is as follows:

[0075] Input data: A standardized dataset X with a dimension of (T, d), where T is the time step and d is the feature dimension.

[0076] Convolution operation: Use one or more 1D convolutional layers to process time series data. The input of each convolutional layer is the output of the previous layer, and the output is a new feature map.

[0077] Convolution kernel size and step size: Choose an appropriate convolution kernel size and step size to ensure that the convolution operation can effectively extract local features. Assume that the convolution kernel size of the lth convolution layer is k l , the step length is s l .

[0078] Activation function: Each convolutional layer is followed by a nonlinear activation function, such as ReLU (Rectified Linear Unit), to introduce nonlinear characteristics.

[0079] The output of the convolutional layer is the convolutional feature map F l , whose dimension is (T′, d l ), where T′ is the number of time steps after the convolution operation, d l is the output feature dimension of the convolutional layer.

[0080] The mathematical description of an LSTM unit is as follows:

[0081] Forget Gate:

[0082] f t =σ(W f ·[h t-1 , x t ]+b f )

[0083] Where σ is the sigmoid activation function, W f is the weight matrix, b f is the bias vector, h t-1 is the hidden state at the previous moment, x t is the current input.

[0084] Input Gate:

[0085] i t =σ(W i ·[h t-1 , x t ]+b i )

[0086]

[0087] Among them, W i and W c is the weight matrix, b i and b c is the bias vector, tanh is the tanh activation function, is the candidate memory cell state.

[0088] Cell State Update:

[0089]

[0090] Among them, C t is the current memory cell state, C t-1 is the state of the memory unit at the previous moment.

[0091] Output Gate:

[0092] o t =σ(W o ·[h t-1 , x t ]+b o )

[0093] h t =o t tanh(C t )

[0094] Among them, W o is the weight matrix, b o is the bias vector, h t is the hidden state at the current moment.

[0095] Construct a two-layer LSTM network, each containing 64 LSTM units, followed by a hidden layer using the ReLU activation function, and finally an output layer using a linear activation function.

[0096] Assume that the input sequence length is T and the feature dimension is d, then the input matrix X dimension is (T, d).

[0097] First layer LSTM:

[0098] H 1 =LSTM 1 (X)

[0099] Among them, H 1 is the output of the first layer of LSTM, with a dimension of (T, 64).

[0100] Second layer LSTM:

[0101] H 2 =LSTM 2 (H 1 )

[0102] Among them, H 2 is the output of the second layer LSTM, with a dimension of (T, 64).

[0103] Hidden Layer:

[0104] H hidden =ReLU(W hidden ·H 2 +b hidden )

[0105] Among them, W hidden is the weight matrix, b hidden is the bias vector, and the output dimension is (T, 32).

[0106] Output layer:

[0107] Y output =W output ·H hidden +b output

[0108] Among them, W output is the weight matrix, b output is the bias vector, and the output dimension is (T, 1).

[0109] The mean square error (MSE) is used as the loss function, which is defined as:

[0110]

[0111] Among them, yi is the true value, is the predicted value and n is the number of samples.

[0112] The Adam optimizer is used for model training. The Adam optimizer combines the advantages of momentum and RMSProp and accelerates gradient descent by adaptively adjusting the learning rate.

[0113] The formula for updating parameters of the Adam optimizer is as follows:

[0114] m t =β1m t-1 +(1-β1)g t

[0115]

[0116]

[0117]

[0118]

[0119] Among them, θ t is the model parameter, g t is the gradient, m t and v t are the first-order moment estimate and the second-order moment estimate of the gradient, β1 and β2 are momentum terms, α is the learning rate, and ∈ is a decimal to prevent division by zero.

[0120] Assume the training data is Where N is the number of training samples. The training process is as follows:

[0121] Step 1: Initialize model parameters θ.

[0122] Step 2: Set the number of training rounds E (here E represents the number of training rounds) and the batch size B.

[0123] Step 3: For each round e (here e represents the current training round number, in the first training round, e=1, in the second training round, e=2, and so on, until e=E):

[0124] 1). Randomly shuffle the training data.

[0125] 2). Divide the training data into several batches, each batch size is B.

[0126] 3). For each batch

[0127] i. Forward propagation calculates the predicted value

[0128] ii. Calculate the loss function MSE.

[0129] iii. Back propagation calculates the gradient g t .

[0130] iv. Update model parameters θ.

[0131] Step 4: Repeat step 3 until the training is completed.

[0132] Through the above steps, the model can effectively learn the time series characteristics of high-voltage users' electricity consumption, and then make accurate electricity consumption forecasts. The performance of the model can be evaluated by the mean square error and graphical methods.

[0133] Error metrics evaluate the accuracy of the model by calculating the difference between the predicted value and the actual value. Commonly used error metrics include mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE). They are calculated as follows:

[0134] Mean Square Error (MSE)

[0135] This method measures the average of the sum of squares of the differences between the predicted values ​​and the actual values, and the formula is:

[0136]

[0137] Among them, y i is the actual value of the ith sample, is the predicted value of the ith sample, and n is the total number of samples.

[0138] Root mean square error (RMSE)

[0139] This method is the square root of the MSE and provides an error measure in the same units as the actual value. The formula is:

[0140]

[0141] Mean absolute error (MAE)

[0142] The mean absolute error (MAE) is the average of the absolute differences between the predicted values ​​and the actual values, and the formula is:

[0143]

[0144] These error metrics can be used to quantitatively evaluate the accuracy of model predictions.

[0145] In addition to using methods such as mean square error to calculate error metrics to evaluate the quality of the model, graphical methods can also be used to intuitively display the relationship between the prediction results and the actual data to better understand and analyze the performance of the model. Commonly used graphical methods include time series graphs and residual graphs.

[0146] The time series graph shows the comparison between actual values ​​and predicted values ​​over time. Through intuitive comparison, the prediction performance of the model can be evaluated.

[0147] Assume that the actual value time series is The predicted value time series is The method of drawing a time series graph is as follows:

[0148] 1) With time t as the horizontal axis, the actual value y t and predicted values is the vertical coordinate;

[0149] 2) Draw the actual value time series Line chart of

[0150] 3) Draw the predicted value time series Line chart of .

[0151] By comparing these two broken lines, you can intuitively see the prediction effect of the model.

[0152] The residual plot shows the prediction error (residual) at each time point, that is, the difference between the predicted value and the actual value. The residual plot helps to examine the distribution and characteristics of the model error.

[0153] Assume the residual sequence is in The method of drawing the residual graph is as follows:

[0154] 1) With time t as the horizontal axis, the residual e t Is the vertical axis.

[0155] 2) Draw the residual sequence A scatter plot or line chart.

[0156] Residual plots can be used to detect whether there are systematic forecast biases (such as whether the residuals show a certain trend) and whether the distribution of errors is uniform.

[0157] The present invention proposes a method for predicting power consumption of high-voltage users based on a long short-term memory (LSTM) neural network. The power consumption data of high-voltage users is taken as the research object. By cleaning the daily power consumption data of a single high-voltage user, the influence of outlier data on the performance of the prediction model is reduced. Through data classification and feature engineering, the original data is converted into a more representative feature set to facilitate model learning. The volatility of the data is reduced by data normalization, thereby improving the stability and convergence speed of model training. Then a CNN-LSTM hybrid model is constructed. One or more convolutional layers are added before the LSTM network to extract local features in the data, and then a two-layer LSTM neural network model is constructed, each layer containing 64 LSTM units, and a model that can effectively learn the time series characteristics of power consumption of high-voltage users and then accurately predict power consumption is trained. After completing the construction and training of the model, the present invention also effectively and intuitively evaluates the prediction results of the model through mean square error and graphical methods.

[0158] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for predicting high-voltage user electricity consumption data based on CNN-LSTM, characterized by: The following steps are included: S1: Data processing of historical electricity consumption data of high-voltage users, including data cleaning, data classification and feature engineering, data normalization, data splitting and sequence creation; S2: Use the processed high-voltage user data to build a CNN-LSTM hybrid model; add a convolution layer before the LSTM network to extract local features in the data, and then build a two-layer LSTM neural network model, each layer contains 64 LSTM units; and use the training set data to train the model; S3: Use error metrics and graphical methods to evaluate the prediction effect of the model; put the evaluated model into use and input data to obtain the corresponding prediction results.

2. According to a method for predicting high-voltage user electricity consumption data based on CNN-LSTM according to claim 1, it is characterized in that: Step S1 specifically comprises: S11: Data cleaning: Use the local outlier factor (LOF) method to clean the historical power consumption data of high-voltage users and remove unreasonable outlier data; S12: Data classification and feature engineering: classify and accumulate data according to the region, sector, industry, and group to which the high-voltage users belong, further decompose date features into information of different time granularities such as month, week, and day, and perform feature engineering processing; S13: Data normalization, using Z-score standardization to normalize each feature to reduce data volatility and ensure the stability and convergence of model training; S14: Data splitting and sequence creation, split the data into training and test sets, and convert them into a time series format suitable for LSTM model input.

3. According to a CNN-LSTM-based high-voltage user electricity consumption data prediction method according to claim 2, it is characterized by: The data cleaning process of step S11 is specifically as follows: S111: For each data point x in the data set i , calculate its difference with all other data points x j The Euclidean distance of: d(x i ,x j )=||x i -x j || S112: For data point x i , find its k nearest neighbor data points, called k-neighborhood, denoted by N k (x i ); S113: Define data point x j For data point x i The reachable distance is: reach-dist k (x i ,x j )=max{d(x i ,x j ),k-distance(x j )} Among them, k-distance(x j ) is the kth nearest neighbor to x j distance; S114: Define data point x i The local reachable density is: Among them, N k (x i ) is the data point x i k-neighborhood S115: Finally, define the data point x i The local outlier factor of is: When LOF k (x i ) value is much greater than 1, the data point x i are considered outliers.

4. The method for predicting high-voltage user electricity consumption data based on CNN-LSTM according to claim 2 is characterized in that: The data classification and feature engineering of step S12 are specifically as follows: S121: Classify the data according to the region, sector, industry, and group to which the high-voltage users belong; suppose that the data set D contains n data points, each of which has m features; divide the data set into C subsets {D1, D2, ..., D C }; S122: Decompose the date feature into month, week, and day information; assuming the date feature is t, the following features are generated: month m = month(t), week w = dayofweek(t), day d = day(t); S123: Further combine the original data and the generated date features to enhance the feature expression ability of the model and complete data classification; combine the features f new =f i ×f j ; Aggregate feature f agg =∑ k f k .

5. The method for predicting high-voltage user electricity consumption data based on CNN-LSTM according to claim 2 is characterized in that: The data splitting and sequence creation in step S14 are specifically as follows: S141: Divide the standardized data set into a training set D according to the proportion p train and the test set D test ; S142: Assume that the time step is T, and convert the data set into a time series format; each time series sample contains data of T time steps; define a time series sample S i for: S i ={V i ,V i+1 ,...,V i+T-1 } Among them, V i is the eigenvector at time step i.

6. The method for predicting high-voltage user electricity consumption data based on CNN-LSTM according to claim 1 is characterized in that: Step S2 specifically includes: S21: Build a CNN-LSTM hybrid model, add one or more convolutional layers before the LSTM network to extract local features in the data, and then build a two-layer LSTM neural network model, each layer contains 64 LSTM units, followed by a hidden layer using an activation function, and finally an output layer using a linear activation function; The mean square error MSE is used as the loss function; The Adam optimizer is used for model training; the Adam optimizer combines the advantages of momentum and RMSProp and accelerates gradient descent by adaptively adjusting the learning rate; S22: Model training process, design training data samples; initialize model parameters; set the number of training rounds and batch size; repeat the following four steps for each training until the model training is completed; 1) Forward propagation calculates the predicted value; 2) Calculate the loss function; 3) Back propagation calculates gradients; 4) Update model parameters.

7. The method for predicting high-voltage user electricity consumption data based on CNN-LSTM according to claim 6 is characterized in that: The model training process of step S22 is specifically as follows: Assume the training data is Where N is the number of training samples; the training process is as follows: S221: Initialize model parameters θ; S222: Set the number of training rounds E (where E represents the number of training rounds) and the batch size B; S223: For each round e (here e represents the current training round number, in the first training round, e=1, in the second training round, e=2, and so on, until e=E): 1). Randomly shuffle the training data; 2) Divide the training data into different batches, each batch size is B; 3). For each batch i. Forward propagation calculates the predicted value ii. Calculate the loss function MSE; iii. Back propagation calculates the gradient gt; iv. Update model parameters θ; S224: Repeat step S223 until the training is completed.

8. The method for predicting high-voltage user electricity consumption data based on CNN-LSTM according to claim 1, characterized in that: Step S3 specifically comprises: S31: Calculate the error metrics of mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) of the predicted data and sample data to evaluate the quality of the model; S32: Use graphical methods such as time series plots and residual plots to intuitively display the relationship between the prediction results and the actual data to understand and analyze the performance of the model.

Images