CN115688993A - Short-term power load prediction method suitable for power distribution station area - Google Patents

Abstract

The invention discloses a short-term power load prediction method suitable for a distribution station area, which comprises the steps of constructing a one-dimensional Convolutional Neural Network (CNN) and a long-term memory neural network (LSTM) which are respectively used for extracting a local dependency relationship among load data and a long-term change trend of the load data; establishing an LSTNet load prediction model, fusing a traditional autoregressive model to solve the problem that a neural network is insensitive to the extreme value of load data, and using the power load data of a certain distribution area in the training and prediction process of the network; the short-term change trend of the power load is predicted through the established prediction model, and is compared and analyzed with the LSTM, the Bi-LSTM and CNN-LSTM models of the bidirectional long-and-short-term memory neural network, and the LSTNet model is verified to be more advantageous in the aspect of short-term load prediction and higher in prediction accuracy.

Description

A Short-term Power Load Forecasting Method Applicable to Distribution Area

technical field

The invention relates to the technical field of electric power systems, and proposes a short-term power load forecasting method suitable for distribution stations.

Background technique

With the increasing demand for electricity, traditional power grids have begun to encounter challenges in centralized power distribution, manual monitoring and restoration, and two-way communication. The emergence of smart grids provides technical possibilities to solve the above difficulties and helps to monitor power production, transmission, and consumption. , and balance the relationship among the three. However, due to the influence of uncertain factors such as climate, economy, and environment, the power load fluctuates greatly, and it is difficult to predict it simply. In order to ensure the safe operation of the power system, reduce power consumption, meet market demand, balance the supply and demand of power loads, and maximize economic benefits, it is necessary to accurately predict power loads.

Power load forecasting plays an important role in modern power system research. It is not only the prerequisite for ensuring the safe and economical operation of the power system, but also the basis for rationally arranging power grid dispatching plans. According to the forecast duration, load forecasting can be divided into short-term, medium- and long-term, and long-term. Short-term load forecasting is to predict the power load from a few minutes to a week in the future. Accurate short-term load forecasting will help to formulate a reasonable power production plan, avoid excessive waste of power resources, and improve the economic benefits of the power system.

According to the time span, short-term power load forecasting can be divided into day-ahead load forecasting and week-ahead load forecasting, which can provide relevant references for power load scheduling. With the rapid development of power system monitoring technology and information communication technology, more and more power systems have installed high-precision data measurement devices, and the massive data collected has laid a data foundation for accurate short-term power load forecasting.

Load forecasting methods are mainly divided into two categories: statistical methods and machine learning methods. Statistical methods mainly include autoregressive model, Kalman filter method, multiple linear regression method and so on. Statistical methods have the advantages of simple principle and rapid modeling, but when the sample data is too large, its prediction effect is mediocre. Machine learning methods mainly include gray theory, neural network, support vector regression and so on. The machine learning method has the advantages of processing massive data and high prediction accuracy, but it is very dependent on the accuracy of the collected data. Compared with a single deep learning method, the combined deep learning method formed by the fusion of multiple methods is more predictive and universal, and the prediction accuracy is higher.

Contents of the invention

Based on the above background, the present invention proposes to use the LSTNet model for short-term load forecasting. This method can effectively use the local dependencies between load data and the periodic characteristics of long-term changes in load data for forecasting. On this basis, the autoregressive model is added to solve the insensitivity of the neural network to the extreme value of the load data, and further improve the prediction effect. Compared with LSTM, Bi-LSTM and CNN-LSTM algorithms, the LSTNet model has a better effect on short-term load forecasting in distribution stations.

In order to achieve the above object, the technical solution of the present invention is:

A short-term power load forecasting method suitable for distribution stations, comprising the following steps:

S1: Construct a convolutional neural network CNN and a long-short-term memory neural network LSTM to extract the local dependencies between load data and the long-term change trend of load data;

S2: Based on step S1, establish the LSTNet load forecasting model, integrate the traditional autoregressive model to solve the insensitivity of the neural network to the extreme value of the load data, and use the power load data of a certain distribution station area for the training and forecasting process of the network;

S3: Use the prediction model established in step S2 to predict the short-term change trend of electric load, and compare and analyze it with LSTM, bidirectional long-short-term memory neural network (Bi-LSTM) and CNN-LSTM models to verify the LSTNet model in short-term load forecasting It has more advantages and higher prediction accuracy.

Further, in the step S1, the study of the one-dimensional convolutional neural network CNN and the long-short-term memory neural network LSTM is used to extract the local dependencies between the load data and the long-term change trend of the load data, including the following steps:

S1-1: Convolutional neural network (CNN), as one of the typical representative algorithms of deep learning, is a feed-forward neural network including convolution operation and deep analytical structure. CNN is constructed according to the biological visual mechanism. The convolution kernel parameter sharing mechanism of the hidden layer and the sparse characteristics of the interlayer connection greatly reduce the complexity of the network and the parameters that need to be adjusted, avoiding the risk of network overfitting .

What the present invention uses is a zero-filled one-dimensional CNN without a pooling layer, which is mainly used to extract the abstract features of the original data. Assume that the input time series data has a row and b column, and the convolution kernel is m row and n column. Since the convolution kernel only slides in one dimension, and the convolution kernel contains all eigenvectors, the width of the convolution kernel must be consistent with the width of the input time series data, that is, n=b, then the i-th convolution kernel c _i acts On the input matrix X can be expressed as:

a _i =c _i *X+b _i (1)

In the formula, * represents the convolution operation, and _bi is the convolution bias item.

S1-2: Long short-term memory network (LSTM) is a variant of recurrent neural network (RNN), which alleviates the problem of gradient disappearance when RNN is dealing with relatively long-term series data to a certain extent, but when dealing with ultra-long-term series data When , LSTM still has the problem of gradient disappearance, and the corresponding solution will be given in the following steps.

The core of LSTM is the gating unit, there are three main types: forget gate, update gate and output gate. Among them, the forget gate selectively forgets part of the historical data, the update gate fuses the current retained information with the historical legacy data, and the output gate measures the influence of the current state on the hidden layer. The calculation process of LSTM is as follows:

c^t分别为t时刻更新门、遗忘门、输出门和记忆细胞的状态，W_u、W_n、W_o、W_c和b_u、b_n、b_o、b_c分别为对应的权值矩阵和偏置项，σ、tanh代表激活函数。In the formula,

c ^t is the state of update gate, forget gate, output gate and memory cell at time t respectively, Wu _u , W _n , W _o , W _c and b _u , b _n , b _o , b _c are the corresponding weight matrix And the bias term, σ, tanh represent the activation function.

σ＝1/(1+e ^-x ) (8)

tanh＝(e ^x -e ^-x )/(e ^x +e ^-x ) (9)

Further, in the step S2, based on the step S1, the LSTNet load forecasting model is established, and the traditional autoregressive model is integrated to solve the problem of the insensitivity of the neural network to the extreme value of the load data, and the power load data of a certain distribution station area is used for The training and prediction process of the network includes the following steps:

S2-1: The data in real life are generally incomplete and missing, and cannot be directly used for data mining, or the mining effect is poor. In order to improve the quality of data mining, data preprocessing technology is produced. Data preprocessing refers to the necessary processing methods such as review, screening, sorting, etc. before the collected data is analyzed.

There are many methods of data preprocessing: data cleaning, data integration, data transformation, data specification, etc. Data normalization processing is a commonly used data preprocessing method. The present invention uses the most value normalization to preprocess data, and its formula is as follows:

x _scale =(xx _min )/(x _max -x _min ) (10)

In the formula, x is the measured data, and x _min and x _max are the minimum and maximum values of the measured data.

S2-2: The LSTNet model consists of a linear part and a nonlinear part. The linear part contains the periodic components in the time series data, and the nonlinear part focuses on the local extreme cases of the time series data.

The first layer of the nonlinear part is a zero-filled one-dimensional CNN without a pooling layer, and is activated using the ReLU activation function. Then the i-th convolution kernel _ci acting on the input matrix X can be expressed as:

a _i ＝ReLU(c _i *X+b _i ) (11)

In the formula, ReLU(x)=max(0,x), * indicates the convolution operation, and b _i is the convolution bias item.

In order to solve the problem of gradient disappearance that occurs when LSTM processes ultra-long sequence data, the present invention adds an LSTM network component connected by time jumps to the LSTNet model. This component utilizes the periodic characteristics of the power load. For the electricity load at one o'clock, it is more meaningful to refer to the electricity load data at one o'clock in the afternoon in the historical data than to refer to the load data at noon and morning of this day. Then the output results of the convolution are input to the LSTM component and the skip-connected LSTM component respectively. The specific update process is as follows:

和t-p+1时刻到t时刻跳过的隐态

的加权组合，全连接层的输出可表示为：At the end of the non-linear part, a fully connected layer is used to combine the outputs of both. The input of the fully connected layer is the hidden state of the LSTM component at time t

and the hidden state skipped from time t-p+1 to time t

The weighted combination of , the output of the fully connected layer can be expressed as:

为非线性部分的输出，W_LSTM与

为权重，b_D为偏置项。In the formula,

is the output of the nonlinear part, W _LSTM and

is the weight, and b _D is the bias term.

In addition, due to the nonlinear nature of CNN and LSTM, the neural network is not sensitive to extreme values of input data, which will have a great impact on the prediction accuracy of the model. Therefore, a linear regression part is added to the LSTNet model, and the linear regression part is composed of an autoregressive model. The process is as follows:

与x^t-i表示线性部分的输出与输入，W_i与b_L分别表示权重与偏置，n为输入窗口的大小。In the formula,

and x ^ti represent the output and input of the linear part, W _i and b _L represent the weight and bias, respectively, and n is the size of the input window.

The output of the LSTNet model is the result y ^t obtained after adding the linear and nonlinear output results and then passing through the Sigmoid activation function, which can be expressed by the following formula.

S2-3: the present invention adopts root mean square error (RMSE), average absolute percentage error (MAPE) and R2_score as the evaluation index of LSTNet prediction model, and the calculation formula is as follows:

表示模型的预测值，y_i为真实值，

为真实值的平均值

当RMSE＝0，MAPE＝0，R2_score＝1时，模型拟合效果最好。in,

Indicates the predicted value of the model, y _i is the real value,

is the average of the true values

When RMSE=0, MAPE=0, R2_score=1, the model fitting effect is the best.

In said step S3, the short-term variation trend of electric load is predicted by the prediction model established in step S2, and comparative analysis is carried out with LSTM, bidirectional long-short-term memory neural network (Bi-LSTM) and CNN-LSTM model, verify LSTNet The model has more advantages and higher prediction accuracy in short-term load forecasting, including the following steps:

S3-1: Use the load data recorded on a public transformer, that is, the power load data in the distribution station area;

S3-2: In order to highlight the advantages of the LSTNet model in short-term load forecasting, LSTM, Bi-LSTM, and CNN-LSTM forecasting models were built for comparison. These three models and the LSTNet model all use the data set mentioned above and perform training;

S3-3: In order to verify the effect of each component in the LSTNet model on the prediction results, the prediction results of the LSTNet model and the prediction results of the ablation model were compared and analyzed on the time scale of one day and one week. Based on the LSTNet model, the model after removing the linear autoregressive component is the LSTNetW/OAR model, the model after removing the LSTM component of the intermediate layer skip connection is the LSTNetW/OSkip model, and the model after removing the linear autoregressive component and the loop skip component is Basic LSTM model.

The beneficial effects of the present invention are:

1) The prediction accuracy of the LSTNet model is higher than that of LSTM, Bi-LSTM, and CNN-LSTM, and it is more suitable for short-term load prediction in distribution stations;

2) Considering the cycle jumping component and the autoregressive model as the components of the LSTNet model, it can avoid the problem of insensitivity to the extreme value of the load data while improving the prediction accuracy;

3) Compared with a single deep learning method, the combined deep learning method is more universal in forecasting problems, and can be used in forecasting research in other fields such as photovoltaics, wind power, and comprehensive energy.

Description of drawings

Fig. 1 is a one-dimensional CNN calculation process of the present invention;

Fig. 2 is the basic structure of LSTM of the present invention;

Fig. 3 is the basic structure of the LSTNet model of the present invention;

Fig. 4 is the LSTNet model load forecasting process of the present invention;

Fig. 5 is the electric load annual variation curve of the present invention;

Fig. 6 is the one-day prediction result of electric load of the present invention;

Fig. 7 is the electric load one day prediction error of the present invention;

Fig. 8 is the prediction result of electric load of the present invention for one week;

Fig. 9 is the electric load one week forecast error of the present invention;

Figure 10 is the one-day prediction result of the ablation experiment of the present invention;

Fig. 11 is the one-day prediction error of the ablation experiment of the present invention;

Fig. 12 is the prediction result of one week of the ablation experiment of the present invention;

Fig. 13 is the prediction error of one week of the ablation experiment of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Referring to Figures 1 to 13, a short-term power load forecasting method suitable for distribution stations includes the following steps:

S1: Construct a one-dimensional convolutional neural network CNN and a long-short-term memory neural network LSTM to extract the local dependencies between load data and the long-term change trend of load data;

S2: Based on step S1, establish the LSTNet load forecasting model, integrate the traditional autoregressive model to solve the insensitivity of the neural network to the extreme value of the load data, and use the power load data of a certain distribution station area for the training and forecasting process of the network;

S3: Use the prediction model established in step S2 to predict the short-term change trend of electric load, and compare and analyze it with LSTM, bidirectional long-short-term memory neural network (Bi-LSTM) and CNN-LSTM models to verify the LSTNet model in short-term load forecasting It has more advantages and higher prediction accuracy.

Further, in the step S1, the study of the one-dimensional convolutional neural network CNN and the long-short-term memory neural network LSTM is used to extract the local dependencies between the load data and the long-term change trend of the load data, including the following steps:

S1-1: Convolutional neural network (CNN), as one of the typical representative algorithms of deep learning, is a feed-forward neural network including convolution operation and deep analytical structure. CNN is constructed according to the biological visual mechanism. The convolution kernel parameter sharing mechanism of the hidden layer and the sparse characteristics of the interlayer connection greatly reduce the complexity of the network and the parameters that need to be adjusted, avoiding the risk of network overfitting .

What the present invention uses is a zero-filled one-dimensional CNN without a pooling layer, which is mainly used to extract abstract features of raw data, as shown in FIG. 1 . Assume that the input time series data has a row and b column, and the convolution kernel is m row and n column. Since the convolution kernel only slides in one dimension, and the convolution kernel contains all eigenvectors, the width of the convolution kernel must be consistent with the width of the input time series data, that is, n=b, then the i-th convolution kernel c _i acts On the input matrix X can be expressed as:

a _i =c _i *X+b _i (1)

In the formula, * represents the convolution operation, and _bi is the convolution bias item.

S1-2: Long short-term memory network (LSTM) is a variant of recurrent neural network (RNN), which alleviates the problem of gradient disappearance when RNN is dealing with relatively long-term series data to a certain extent, but when dealing with ultra-long-term series data When , LSTM still has the problem of gradient disappearance, and the corresponding solution will be given below. Figure 2 shows the basic structure of LSTM.

The core of LSTM is the gating unit, there are three main types: forget gate, update gate and output gate. Among them, the forget gate selectively forgets part of the historical data, the update gate fuses the current retained information with the historical legacy data, and the output gate measures the influence of the current state on the hidden layer. The calculation process of LSTM is as follows:

c^t分别为t时刻更新门、遗忘门、输出门和记忆细胞的状态，W_u、W_n、W_o、W_c和b_u、b_n、b_o、b_c分别为对应的权值矩阵和偏置项，σ、tanh代表激活函数。In the formula,

c ^t is the state of update gate, forget gate, output gate and memory cell at time t respectively, Wu _u , W _n , W _o , W _c and b _u , b _n , b _o , b _c are the corresponding weight matrix And the bias term, σ, tanh represent the activation function.

σ＝1/(1+e ^-x ) (8)

tanh＝(e ^x -e ^-x )/(e ^x +e ^-x ) (9)

Further, in the step S2, based on the step S1, the LSTNet load forecasting model is established, and the traditional autoregressive model is integrated to solve the problem of the insensitivity of the neural network to the extreme value of the load data, and the power load data of a certain distribution station area is used for The training and prediction process of the network includes the following steps:

S2-1: The data in real life are generally incomplete and missing, and cannot be directly used for data mining, or the mining effect is poor. In order to improve the quality of data mining, data preprocessing technology is produced. Data preprocessing refers to the necessary processing methods such as review, screening, sorting, etc. before the collected data is analyzed.

There are many methods of data preprocessing: data cleaning, data integration, data transformation, data specification, etc. Data normalization processing is a commonly used data preprocessing method. The present invention uses the most value normalization to preprocess data, and its formula is as follows:

x _scale =(xx _min )/(x _max -x _min ) (10)

In the formula, x is the measured data, and x _min and x _max are the minimum and maximum values of the measured data.

S2-2: The LSTNet model consists of a linear part and a nonlinear part. The linear part contains the periodic components in the time series data, and the nonlinear part focuses on the local extreme cases of the time series data. The basic structure of the LSTNet model is shown in Figure 3.

The first layer of the nonlinear part is a zero-filled one-dimensional CNN without a pooling layer, and is activated using the ReLU activation function. Then the i-th convolution kernel _ci acting on the input matrix X can be expressed as:

a _i ＝ReLU(c _i *X+b _i ) (11)

In the formula, ReLU(x)=max(0,x), * indicates the convolution operation, and b _i is the convolution bias item.

In order to solve the problem of gradient disappearance that occurs when LSTM processes ultra-long sequence data, the present invention adds an LSTM network component connected by time jumps to the LSTNet model. This component utilizes the periodic characteristics of the power load. For the electricity load at one o'clock, it is more meaningful to refer to the electricity load data at one o'clock in the afternoon in the historical data than to refer to the load data at noon and morning of this day. Then the output results of the convolution are input to the LSTM component and the skip-connected LSTM component respectively. The specific update process is as follows:

和t-p+1时刻到t时刻跳过的隐态

的加权组合，全连接层的输出可表示为：At the end of the non-linear part, a fully connected layer is used to combine the outputs of both. The input of the fully connected layer is the hidden state of the LSTM component at time t

and the hidden state skipped from time t-p+1 to time t

The weighted combination of , the output of the fully connected layer can be expressed as:

为非线性部分的输出，W_LSTM与

为权重，b_D为偏置项。In the formula,

is the output of the nonlinear part, W _LSTM and

is the weight, and b _D is the bias item.

In addition, due to the nonlinear nature of CNN and LSTM, the neural network is not sensitive to extreme values of input data, which will have a great impact on the prediction accuracy of the model. Therefore, a linear regression part is added to the LSTNet model, and the linear regression part is composed of an autoregressive model. The process is as follows:

与x^t-i表示线性部分的输出与输入，W_i与b_L分别表示权重与偏置，n为输入窗口的大小。In the formula,

and x ^ti represent the output and input of the linear part, W _i and b _L represent the weight and bias, respectively, and n is the size of the input window.

The output of the LSTNet model is the result y ^t obtained after adding the linear and nonlinear output results and then passing through the Sigmoid activation function, which can be expressed by the following formula.

S2-3: the present invention adopts root mean square error (RMSE), average absolute percentage error (MAPE) and R2_score as the evaluation index of LSTNet prediction model, and the calculation formula is as follows:

表示模型的预测值，y_i为真实值，

为真实值的平均值

当RMSE＝0，MAPE＝0，R2_score＝1时，模型拟合效果最好。in,

Indicates the predicted value of the model, y _i is the real value,

is the average of the true values

When RMSE=0, MAPE=0, R2_score=1, the model fitting effect is the best.

The established LSTNet model is used in the short-term forecasting process of electric load, and the specific algorithm flow is shown in Figure 4.

In said step S3, the short-term variation trend of electric load is predicted by the prediction model established in step S2, and comparative analysis is carried out with LSTM, bidirectional long-short-term memory neural network (Bi-LSTM) and CNN-LSTM model, verify LSTNet The model has more advantages and higher prediction accuracy in short-term load forecasting, including the following steps:

S3-1: The present invention uses the load data recorded on a public transformer in a community by the State Grid Wenzhou Electric Power Company, that is, the power load data of the distribution station area. The data sampling frequency is 15 minutes, a total of 12 months, and a total of 35,133 records are recorded data. Figure 5 shows the power load change curve for one year, and the small picture shows the load data for the first week of this year.

It can be seen from Figure 5 that the annual change curve of power load in the distribution station area presents strong nonlinearity and seasonality, that is, the power load in summer is higher than that in other seasons, which is consistent with the power load habits in southern my country. In addition, the change curve of power load cycle presents an obvious periodic characteristic, which can be used for short-term load forecasting research in order to formulate a reasonable power grid dispatching plan.

S3-2: In order to highlight the advantages of the LSTNet model in short-term load forecasting, the present invention also builds LSTM, Bi-LSTM, and CNN-LSTM forecasting models for comparison. These three models and the LSTNet model all use the data set mentioned above, and the training times are 100 times. Figure 6, Figure 7, Figure 8, and Figure 9 are the prediction result curves and prediction error curves of the four prediction models on the time scales of one day and one week respectively. Table 1 shows the evaluation indicators of the model accuracy of the four forecasting models on the time scale of one day and one week.

S3-3: In order to verify the effect of each component in the LSTNet model on the prediction results, the present invention compares and analyzes the prediction results of the LSTNet model and its ablation model on the time scale of one day and one week. Based on the LSTNet model, the model after removing the linear autoregressive component is the LSTNetW/OAR model, the model after removing the LSTM component of the intermediate layer skip connection is the LSTNetW/OSkip model, and the model after removing the linear autoregressive component and the loop skip component is Basic LSTM model. Figure 10, Figure 11, Figure 12, and Figure 13 are the prediction result curves and prediction error curves of the LSTNet model and its ablation model on the time scale of one day and one week, respectively. Table 2 shows the model accuracy evaluation indicators of the LSTNet model and its ablation model on the time scale of one day and one week.

In order to make those skilled in the art better understand the present invention, the example analysis includes the following formations:

1. Calculation example description and simulation result analysis

In order to highlight the advantages of the LSTNet model in short-term load forecasting, the present invention builds LSTM, Bi-LSTM, and CNN-LSTM forecasting models for comparison. These three models and the LSTNet model all use the data set mentioned above, and the training times are 100 times. Figure 6, Figure 7, Figure 8, and Figure 9 are the prediction result curves and prediction error curves of the four prediction models on the time scales of one day and one week respectively. Table 1 shows the evaluation indicators of the model accuracy of the four forecasting models on the time scale of one day and one week.

It can be seen from Fig. 6 to Fig. 9 that these four prediction methods can better predict the change of power load in the distribution station area regardless of the time scale of one day or one week, but the result curve of the prediction method of the present invention fits best For the real power load change curve of the distribution station, the prediction results of the other three have large deviations when the load changes greatly. At the same time, the power load forecasting method proposed by the present invention has the smallest error on the time scale of one day and one week, and the variation range of the error curve is also smaller than the other three methods.

Table 1 Evaluation indicators of accuracy of different prediction models

Table 1 Accuracy evaluation index of different forecast models

It can be seen from Table 1 that the MSE, MAPE, and R2_score values of the LSTNet model on a one-day time scale are 12.41, 6.46, and 96.72, respectively. The MSE and MAPE values are smaller than the corresponding values of the CNN-LSTM, Bi-LSTM, and LSTM models, while R2_score The value is greater than the corresponding value of the CNN-LSTM, Bi-LSTM, and LSTM models, indicating that among the four models, the LSTNet model has the highest prediction accuracy and the best prediction effect on the one-day time scale. In addition, the MSE, MAPE, and R2_score values of the LSTNet model on the one-week time scale are 19.01, 9.49, and 92.08, respectively. The MSE and MAPE values are the smallest among the four models, and the R2_score value is the largest among the four models. The model accuracy of the one-week time scale The evaluation index has a similar numerical relationship with the model evaluation index on the one-day time scale, indicating that among the four models, the LSTNet model also achieves the best load forecasting effect on the one-week time scale.

In order to verify the effect of each component in the LSTNet model on the prediction results, the present invention compares and analyzes the prediction results of the LSTNet model and its ablation model on the time scale of one day and one week. Based on the LSTNet model, the model after removing the linear autoregressive component is the LSTNetW/OAR model, the model after removing the LSTM component of the intermediate layer skip connection is the LSTNetW/OSkip model, and the model after removing the linear autoregressive component and the loop skip component is Basic LSTM model. Figure 10, Figure 11, Figure 12, and Figure 13 are the prediction result curves and prediction error curves of the LSTNet model and its ablation model on the time scale of one day and one week, respectively. Table 2 shows the model accuracy evaluation indicators of the LSTNet model and its ablation model on the time scale of one day and one week.

It can be seen from Figure 10 and Figure 11 that the LSTNet model has the best prediction effect on the time scale of one day, followed by the LSTNetW/OAR and LSTNetW/OSkip ablation models, and finally the LSTM model, which embodies the autoregressive component and loop The important role of skipping components in improving prediction accuracy. In addition, the LSTNet and LSTNetW/OSkip models integrated with linear autoregressive components have better performance when the data changes are extreme, which reflects the role of linear autoregressive components in alleviating the insensitivity of neural networks to extreme values of data.

It can be seen from Figure 12 and Figure 13 that the LSTNet model still obtains the best prediction effect on the one-week time scale. Compared with the LSTNetW/OSkip ablation model and the LSTM model, the prediction accuracy of the LSTNet model and the LSTNetW/OAR model is higher. The large improvement reflects that on a large time scale, the loop jumping component has a considerable effect on improving the prediction accuracy.

Table 2 Evaluation indicators of LSTNet and its ablation model accuracy

Table 2 Accuracy evaluation index of LSTNet ablation experiment

It can be seen from Table 2 that the MSE, MAPE, and R2_score values of the LSTNetW/OAR model on the one-day time scale are 19.92, 9.95, and 91.55, respectively, and the MSE, MAPE, and R2_score values of the LSTNetW/OSkip model on the one-day time scale are 21.27, respectively. , 11.31, 90.37. The MSE and MAPE values of the LSTNetW/OAR and LSTNetW/OSkip models are larger than the corresponding values of the LSTNet model, but smaller than the corresponding values of the LSTM model. At the same time, the R2_score values of the LSTNetW/OAR and LSTNetW/OSkip models are smaller than the corresponding values of the LSTNet model. However, they are all greater than the corresponding values of the LSTM model, indicating that the loop jump component and the linear autoregressive component have a significant effect on improving the prediction accuracy of the model.

In summary, compared with the previous prediction models of LSTM, Bi-LSTM and CNN-LSTM, the LSTNet model has more advantages and higher prediction accuracy in short-term load forecasting.

In this specification, the schematic representations of the present invention are not necessarily aimed at the same embodiment or example, and those skilled in the art can combine and combine different embodiments or examples described in this specification. In addition, the content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the implementation cases. The protection scope of the present invention also includes Equivalent technical means that the skilled person can think of based on the concept of the present invention.

Claims (4)

1. A short-term electric load forecasting method applicable to power distribution station area, is characterized in that, described method comprises the following steps: S1: Construct a convolutional neural network CNN and a long-short-term memory neural network LSTM to extract the local dependencies between load data and the long-term change trend of load data; S2: Based on step S1, establish the LSTNet load forecasting model, integrate the traditional autoregressive model to solve the insensitivity of the neural network to the extreme value of the load data, and use the power load data of the distribution station area for the training and forecasting process of the network; S3: Use the forecasting model established in step S2 to predict the short-term change trend of electric load, and compare and analyze it with LSTM, bidirectional long-short-term memory neural network Bi-LSTM and CNN-LSTM models to verify that the LSTNet model is better in short-term load forecasting It has advantages and higher prediction accuracy. 2. A kind of short-term power load forecasting method suitable for distribution station areas as claimed in claim 1, characterized in that, in the step S1, the specific steps are as follows: S1-1: Convolutional neural network CNN uses a zero-filled one-dimensional CNN without pooling layer to extract the abstract features of the original data; suppose the input time series data has a row and b column, and the convolution kernel is m rows and n columns; since the convolution kernel only slides in one dimension, and the convolution kernel contains all feature vectors, the width of the convolution kernel must be consistent with the width of the input time series data, that is, n=b, then the i-th volume The product kernel _ci acting on the input matrix X can be expressed as: a _i =c _i *X+b _i (1) In the formula, * represents the convolution operation, and _bi is the convolution bias item; S1-2: The core of LSTM is the gating unit, including: forget gate, update gate and output gate; among them, the forget gate selectively forgets part of the historical data, the update gate fuses the current retained information with the historical legacy data, and the output gate Measure the impact of the current state on the hidden layer; the calculation process of LSTM is as follows:

In the formula,

c ^t is the state of update gate, forget gate, output gate and memory cell at time t respectively, Wu _u , W _n , W _o , W _c and b _u , b _n , b _o , b _c are the corresponding weight matrix And the bias term, σ, tanh represent the activation function; σ＝1/(1+e ^-x ) (8) tanh=(e ^x -e ^-x )/(e ^x +e ^-x ) (9). 3. A kind of short-term power load forecasting method suitable for distribution station area as claimed in claim 2, it is characterized in that, described step S2, concrete steps are as follows: S2-1: Use the most value normalization to preprocess the data, the formula is as follows: x _scale =(xx _min )/(x _max -x _min ) (10) In the formula, x is the measured data, and x _min and x _max are the minimum and maximum values of the measured data; S2-2: The LSTNet model consists of a linear part and a nonlinear part. The linear part contains the periodic components in the time series data, and the nonlinear part focuses on the local extreme cases of the time series data; The first layer of the nonlinear part is an all-zero-filled one-dimensional CNN without a pooling layer, and is activated using the ReLU activation function. Then the i-th convolution kernel c _i acts on the input matrix X and is expressed as: a _i ＝ReLU(c _i *X+b _i ) (11) In the formula, ReLU(x)=max(0,x), * indicates the convolution operation, and b _i is the convolution bias item; In order to solve the gradient disappearance problem of LSTM when processing ultra-long sequence data, the LSTM network component of time skip connection is added to the LSTNet model. In the LSTM component connected to the LSTM component and the jump, the specific update process is as follows:

At the end of the nonlinear part, a fully connected layer is used to combine the outputs of the two; the input of the fully connected layer is the hidden state of the LSTM component at time t

and the hidden state skipped from time t-p+1 to time t

The weighted combination of , the output of the fully connected layer can be expressed as:

In the formula,

is the output of the nonlinear part, W _LSTM and

is the weight, b _D is the bias term; Due to the nonlinear nature of CNN and LSTM, the neural network is not sensitive to the extreme value of the input data, which will have a great impact on the prediction accuracy of the model; therefore, a linear regression part is added to the LSTNet model, and the linear regression part is composed of autoregressive The model is composed, and its process is as follows:

In the formula,

and x ^ti represent the output and input of the linear part, W _i and b _L represent the weight and bias respectively, and n is the size of the input window; The output of the LSTNet model is the result y ^t obtained after adding the linear and nonlinear output results and then passing through the Sigmoid activation function, which is expressed by the following formula:

S2-3: Root mean square error RMSE, mean absolute percentage error MAPE and R2_score are used as the evaluation indicators of the LSTNet prediction model, and the calculation formula is as follows:

in,

Indicates the predicted value of the model, y _i is the real value,

is the average of the true values

4. A kind of short-term power load forecasting method suitable for distribution station area as claimed in claim 3, it is characterized in that, the specific steps of described step S3 are as follows: S3-1: Use the load data recorded on a public transformer, that is, the power load data in the distribution station area; S3-2: Build LSTM, Bi-LSTM, CNN-LSTM prediction models, compare with the LSTNet model, and use the data set in S3-1 for training; S3-3: Compare and analyze the prediction results of the LSTNet model and the prediction results of the ablation model on the time scale of one day and one week; based on the LSTNet model, the model after removing the linear autoregressive component is the LSTNetW/OAR model, removing the middle The model after the LSTM component of the layer skip connection is the LSTNetW/OSkip model, and the model after removing the linear autoregressive component and the loop skip component is the basic LSTM model.

Images