CN116703644A - A short-term power load forecasting method based on Attention-RNN - Google Patents

Abstract

The invention discloses a short-term power load prediction method based on an Attention-RNN, and relates to the technical field of power system load prediction. The current power load prediction has the problem that nonlinear and non-stationary data cannot be processed; the method comprises the following steps: based on an Attention mechanism, quantifying the implicit time sequence correlation among time nodes in a load time sequence, and extracting cross-correlation characteristics; then, the memory characteristics of the RNN network are used to extract trend features and cycle features implicit in the load long-term sequence based on the RNN, and the time-series dependence of the load sequence is mined. And mining the time correlation and long-term dependence characteristics of the load time sequence from the historical data by using an attribute mechanism and RNN network characteristics to form a short-term power load prediction model based on the attribute-RNN, and carrying out short-term power load prediction according to the short-term power load prediction model. According to the method, the load time sequence characteristics and the external multidimensional influence factors are comprehensively considered, and compared with a traditional prediction method, the accuracy of short-term power load prediction is effectively improved.

Description

A short-term power load forecasting method based on Attention-RNN

technical field

The invention relates to the technical field of power system load forecasting, in particular to an Attention-RNN-based short-term power load forecasting method.

Background technique

Short-term power load forecasting refers to the prediction of the power load of the power grid in the next few hours or days, so that the power company can reasonably plan power production and supply, so as to ensure the stable operation of the power grid and meet the electricity demand of users. Short-term power load forecasting usually adopts statistical and machine learning methods based on historical data, and the technologies involved mainly include data collection, preprocessing, feature extraction, model construction and evaluation, etc.

First of all, forecasting electric load needs to collect a large amount of historical data, including various factors such as weather, holidays, working days/rest days, etc. These data are collected through sensors, weather forecasts, electric meters and other means to form a large data set. Next, preprocess the data, including data cleaning, normalization, denoising, etc., to ensure data quality and reliability.

Feature extraction is one of the key steps in building a forecasting model, and it is usually necessary to extract useful features from data through methods such as time series analysis, frequency domain analysis, and wavelet analysis. For example, features such as the average load, the highest load, the lowest load, and load fluctuations of each day in a week can be extracted from historical data.

Model construction is the core part of short-term power load forecasting, and it is necessary to select an appropriate model and train it. Commonly used models include regression models, time series models, neural network models, etc. Among them, neural network models perform well in short-term power load forecasting, especially deep learning models such as long-term short-term memory networks (RNN), which can automatically learn features and have good predictive capabilities.

Finally, in order to evaluate the accuracy and performance of a predictive model, some metrics are needed to measure the forecast error. For example, the root mean square error (RMSE) can be used to assess the error size of the forecast results, or the mean absolute error (MAPE) can be used to assess the degree of deviation of the forecast results.

Although the current power load forecasting algorithm is relatively mature, there are still some shortcomings, such as the inability to deal with nonlinear problems and the inability to deal with non-stationary data. Therefore, more advanced algorithms are needed to improve the prediction accuracy.

Contents of the invention

The technical problem to be solved and the technical task proposed by the present invention are to improve and improve the existing technical solutions, and provide a short-term power load forecasting method based on Attention-RNN, in order to improve the accuracy of short-term power load forecasting. For this reason, the present invention takes the following technical solutions.

A short-term power load forecasting method based on Attention-RNN, comprising the following steps:

1) Determine the characteristic representation method of multivariate heterogeneous data including temperature and holiday information, define the data set format for short-term load forecasting considering multiple load influencing factors, and construct the basic framework of single-step forecasting and multi-step forecasting for day-ahead power load forecasting ;

2) According to the basic framework of day-ahead power load forecasting formed in step 1), based on the Attention mechanism, quantify the time-series correlation implied between each time node in the load time series, and form the day-ahead power load forecasting Attention module;

3) According to the basic framework of the day-ahead power load forecasting formed in step 1), based on the RNN, the hidden trend features and cycle features in the long-term load sequence are extracted, and the timing dependence of the load sequence is mined to form the RNN module of the day-ahead power load forecasting;

4) Based on the Attention module and RNN module formed in step 2) and step 3), build a serial and parallel model integration framework, and use the Attention mechanism and RNN network characteristics to mine the time correlation and long-term load timing from historical data. Depending on the feature, a short-term power load forecasting model based on Attention-RNN is formed, and short-term power load forecasting is performed according to the short-term power load forecasting model.

Based on RNN, this technical solution can adaptively learn key information in long sequence data by introducing the Attention mechanism, and can more accurately capture changes and trends in the sequence. Compared with the traditional RNN model, the RNN model using the Attention mechanism can better deal with the long-term dependence problem, and can adaptively adjust the weight when modeling historical data to improve the prediction accuracy of the model. Due to the introduction of attention mechanism, key information in time series data can be adaptively captured. This method usually uses a combination of gated recurrent units and an attention mechanism, which can improve the modeling ability of long sequence data and has strong generalization ability. It can better handle long-term series data and perform better when the input characteristics are uncertain or the data is noisy.

As a preferred technical means: in step 1), the original data set is expressed as:

In the formula: X is the input matrix; x _i is the input feature vector of the i-th day, which is composed of the 96-point load feature vector l _i and the external factor feature vector f _i ; l _i ∈ R ⁹⁶ is the load vector of the i-th day, the number of points Depends on the sampling frequency; f _i ∈ R ²⁶ is the external factor feature vector of the i-th day, coded by weekday type d _i ∈ R ⁷ , season type code s _i ∈ R ⁴ , monthly type code m _i ∈ R ¹² , holiday The type code h _i ∈ R ² and the meteorological feature vector n _i are composed, and f _i is expressed as:

f _i ＝[d _i s _i m _i h _i n _i ] (16)

Among them, the time-scale feature encoding method adopts one-hot encoding; after determining the data structure of short-term power load forecasting, the basic framework of single-step forecasting and multi-step forecasting of day-ahead power load forecasting is expressed as:

X _i+1:i+T ＝[x _i+1 x _i+2 ... x _i+T ] ^T (19)

F _{i+T+K:i+T+K+τ} ＝[f _i+T+K f _i+T+K+1 ... f _i+T+K+τ ] ^T (20)

In the formula, is the power load forecasting matrix; X _i+1:i+T is the historical input matrix; F _{i+T+K:i+T+K+τ} is the synchronous feature matrix; T and τ are the historical window width and forecast window width ; K is the number of forecast days in advance; Φ is the short-term power load forecasting model; θ ^* is the parameters obtained from model training.

The original data set uses day, season, month, holidays, and weather for short-term power load forecasting, which is conducive to improving forecasting accuracy, generalization ability and reliability, and is of great significance for practical applications. Specifically:

1. Considering multiple factors: By introducing multiple features, it can more comprehensively reflect the complexity and uncertainty of power load changes, including natural environment, social culture, population flow and other factors. These features can provide more accurate input information, thereby improving the accuracy of prediction.

2. Improve the inhomogeneity of data distribution: In the case of considering multiple features, it can improve the inhomogeneity of data distribution and reduce the prediction error. For example, when encountering important events such as holidays or sudden changes in weather, it can better reflect the actual changes in power loads.

3. Enhance the generalization ability of the model: By considering multiple time scales and different types of features, the adaptability of the model to unknown situations can be improved, and the generalization ability of the model can be enhanced, thus making the model more reliable and robust .

In addition, in this technical solution, the time-scale features of one-hot encoding and the basic framework of single-step forecasting and multi-step forecasting can also improve the flexibility, scalability and accuracy of the model, and are suitable for various short-term power load forecasting scenarios.

Use one-hot encoding for time-scale features to facilitate model processing: In neural networks, features in numerical form are easily misunderstood, and one-hot encoding can convert categorical variables into vector forms for neural network processing. This way, the model can better understand the relationship between different points in time in the time series data. In addition, it can avoid misleading the numerical value of the model: the time scale feature in numerical form may mislead the model, making the model mistakenly think that some time points are more important than others. Using one-hot encoding can eliminate this misleading and ensure that each time point is treated equally.

Using the basic framework of single-step forecasting and multi-step forecasting can improve flexibility: the framework of single-step forecasting and multi-step forecasting can be flexibly applied to different data sets and tasks to meet different application requirements. For example, for more detailed prediction tasks, a multi-step prediction framework can be used to improve prediction accuracy and stability; for simple prediction tasks, a single-step prediction framework can be selected to reduce the amount of calculation. Improve scalability: The framework based on single-step forecasting and multi-step forecasting can be combined with other model structures to achieve more complex forecasting tasks. For example, these frameworks can be combined with deep learning methods, statistical models, traditional time series forecasting methods to improve forecasting accuracy and robustness.

As an optimal technical means: in step 2), based on the Attention mechanism, quantify the implicit time series correlation between each time node in the load time series, and extract the cross-correlation features; in order to quantify the time correlation between each time step of the load time series, use "Key-value"mechanism; the input feature vector is transformed into query vector q _i , key vector _ki and value vector v _i through linear mapping; the correlation score a between different time steps is obtained by calculating the dot product of query vector and key vector :

a=q·k (21)

The time correlation between different time steps is obtained by calculating the dot product of the query vector and the key vector at different time steps. The overall process of quantifying the time series correlation of loads based on the Attention mechanism is expressed as:

q _i =W ^q x _i , _ki =W ^k x _i ,v _i =W ^v x _i ,a _i,j =q _i k _j (22)

A＝softmax(K ^T Q), H＝VA (24)

In the formula, W ^q , W ^k , W ^v are the query mapping matrix, key mapping matrix and value mapping matrix; a _{i, j} are the correlation scores between the input vectors x _i and x _j ; A matrix consists of a _{i, j} Composition; K and Q are the key matrix and the query matrix, which are composed of ki _and q _i respectively; H is the Attention output.

The "key-value" mechanism is used to quantify the time correlation between each time step of the load time series, which has many advantages such as clear time series relationship, strong flexibility, and convenient data processing. Clarify the timing relationship: The "key-value" mechanism can clearly describe the relationship between each time step, that is, each "key" represents the characteristics of a time step, and the "value" represents the specific value or vector of the time step. This mechanism enables the model to better grasp the dependencies between different time steps in the time-series data, thus predicting future trends more accurately. Strong flexibility: When using the "key-value" mechanism, you can customize the "key" of each time step according to actual needs, such as using different time scales such as seasons and months, and you can also customize multiple time scales for different application scenarios. A "key" to further improve the prediction accuracy and generalization ability. Convenient data processing: Using the "key-value" mechanism can store and process different information in the same format, which facilitates batch reading and processing of data, and facilitates the combination with the attention mechanism to further improve the performance of the model.

As an optimal technical means: in step 3), the processing formula of the RNN module for power load forecasting based on the RNN-based load time-series dependent feature is shown in (11)-(14);

In the formula, is the weight matrix of the RNN layer l;/> is the bias of the first layer of RNN; σ is the sigmoid activation function; W _o is the linear mapping matrix; /> The hidden state features at time t extracted for the l-layer RNN network; /> is the hidden state feature of the RNN output layer and the moment to be predicted; h _RNN is the feature vector extracted by RNN; L is the number of RNN network layers; T is the width of the RNN input window.

As a preferred technical means: in step 4): the original data set is processed through a sliding window to generate a historical window and a forecast window;

In the serial integration framework: firstly, input the data in the history window into the RNN module, process it through the RNN module, and extract the time series dependent features of the load sequence; then, extract the feature vector group obtained by the RNN module Input the Attention module to extract the correlation between each time step. Finally, input H and the synchronous features in the prediction window into the fully connected layer, and output the short-term power load prediction value;

In the parallel integration framework: first, input the data in the history window into the RNN module and the Attention module in parallel, and mine the time correlation and long-term dependence characteristics of the load sequence; then, the feature vector group extracted by the RNN The Attention output H and the synchronous features in the prediction window are input into the fully connected layer, and the output is the short-term power load forecast value.

Using the history window as input and performing multi-model fusion in a serial ensemble framework or a parallel ensemble framework can fully exploit the complementarity between different models, which is beneficial to improve the prediction accuracy.

Beneficial effects: This technical solution adopts the Attention module and the RNN module. The two modules adopt the serial and parallel model integration framework. The serial integration framework can improve the mapping ability of the prediction model, and reduce the load prediction error by reducing the model prediction variance. The parallel integration framework It can improve the capacity of the forecasting model and reduce the load forecasting error by reducing the model deviation. Based on the Attention mechanism and RNN network characteristics, considering the load timing characteristics and external multi-dimensional influencing factors, the RNN model using the Attention mechanism can better deal with long-term dependence problems, and can adaptively adjust the weight when modeling historical data to improve The predictive accuracy of the model. Due to the introduction of attention mechanism, key information in time series data can be adaptively captured. Compared with traditional forecasting schemes, such as short-term load forecasting methods based on models such as feedforward neural network, convolutional neural network, and recurrent neural network, this technical scheme effectively improves the accuracy of short-term power load forecasting.

Description of drawings

Fig. 1 is a schematic diagram of the overall process of the present invention.

Fig. 2 is a schematic diagram of sliding window processing in the present invention.

Fig. 3 is a schematic diagram of the Attention mechanism of the present invention.

Fig. 4 is a schematic diagram of the RNN of the present invention.

Fig. 5 is a schematic diagram of the serial integration framework of the present invention.

Fig. 6 is a schematic diagram of the parallel integration framework of the present invention.

Fig. 7 is a short-term load forecast curve diagram of the present invention.

Detailed ways

In order to better understand the purpose, technical solution and technical effect of the present invention, the present invention will be further explained below in conjunction with the accompanying drawings.

This embodiment proposes a short-term power load forecasting method based on Attention-RNN, and its implementation process includes the following detailed steps:

Step 1: Propose a characteristic representation method for multiple heterogeneous data such as temperature and holiday information, define the data set format for short-term load forecasting considering multiple load factors, and construct the basic framework of single-step forecasting and multi-step forecasting for day-ahead power load forecasting ;

The original data set of short-term power load forecasting can be expressed as

In the formula: X is the input matrix; x _i is the input feature vector of the i-th day, which is composed of the 96-point load feature vector l _i and the external factor feature vector f _i ; l _i ∈ R ⁹⁶ is the load vector of the i-th day, the number of points Depends on the sampling frequency; f _i ∈ R ²⁶ is the external factor feature vector of the i-th day, coded by weekday type d _i ∈ R ⁷ , season type code s _i ∈ R ⁴ , monthly type code m _i ∈ R ¹² , holiday type coding h _i ∈ R ² and meteorological feature vector n _i , f _i can be expressed as

f _i ＝[d _i s _i m _i h _i n _i ] (2)

Among them, one-hot coding is adopted for time-scale feature coding (weekday type coding, season type coding, etc.), that is, the dimension of the type of the day is set to 1, and the other dimensions are set to 0. After determining the data structure of short-term power load forecasting, the basic framework of single-step forecasting and multi-step forecasting of day-ahead power load forecasting can be expressed as

In the formula, is the power load forecasting matrix; X _i+1:i+T is the historical input matrix; F _{i+T+K:i+T+K+τ} is the synchronous feature matrix; T and τ are the historical window width and forecast window width ; K is the number of forecast days in advance; Φ is the short-term power load forecasting model; θ ^* is the parameters obtained from model training.

The key to short-term power load forecasting modeling is to construct a suitable model and determine the parameters of the model. The parameters of the model in this embodiment are determined by minimizing the expected risk R _exp :

In the formula, represents the loss function. The expected parameters of the final model satisfy:

In practical applications, we approximate the expected risk by calculating the empirical risk. According to the law of large numbers, when there are enough samples, the empirical risk will approach the expected risk:

In order to realize the rolling forecast of the short-term power load, this embodiment proposes a sliding window processing method, as shown in FIG. 2 . The historical load data and synchronous correlation data in the sliding window are used as the input of the short-term power load forecasting model to consider the autocorrelation characteristics of load changes and the cross-correlation characteristics of load and its influencing factors. As the sliding window slides, the forecasting model gradually outputs the load forecasting value.

Step 2: According to the basic framework of the day-ahead power load forecasting formed in step 1, based on the Attention mechanism, quantify the time-series correlation implied between each time node in the load time series, and form the Attention module of the day-ahead power load forecasting;

According to the basic framework of the day-ahead power load forecasting formed in step 1, based on the Attention mechanism, the implicit timing correlation between each time node in the load time series is quantified, and the cross-correlation features are extracted. The calculation process of Attention is shown in Figure 3. In order to quantify the time correlation between each time step of the load time series, a "key-value" mechanism is introduced. The input feature vector is transformed into query vector q _i , key vector _ki and value vector v _i through linear mapping. The correlation score a between different time steps is obtained by calculating the dot product between the key-value vectors:

a=q·k (10)

The time correlation between different time steps is obtained by calculating the dot product between key-value vectors at different time steps. The overall process of quantifying the time series correlation of loads based on the Attention mechanism can be expressed as

q _i =W ^q x _i , _ki =W ^k x _i ,v _i =W ^v x _i ,a _i,j =q _i k _j (11)

A＝softmax(K ^T Q), H＝VA (13)

In the formula, W ^q , W ^k , W ^v are the query mapping matrix, key mapping matrix and value mapping matrix; a _{i, j} are the correlation scores between the input vectors x _i and x _j ; A matrix consists of a _{i, j} Composition; K and Q are the key matrix and query matrix, which are composed of ki _and q _i respectively; H is the output obtained by weighting and summing the value vectors of each time step with the Attention score weight.

Step 3: According to the basic framework of the day-ahead power load forecasting formed in step 1, based on the RNN to extract the implicit trend features and cycle features in the long-term load sequence, mining the time series dependence of the load sequence, and forming the RNN module of the day-ahead power load forecasting;

The processing flow of the load timing dependent feature extraction module based on RNN is shown in (14)-(17).

In the formula, is the weight matrix of the RNN layer l;/> is the bias of the first layer of RNN; σ is the sigmoid activation function; W _o is the linear mapping matrix; /> The hidden state features at time t extracted for the l-layer RNN network; /> It is the hidden state feature of the RNN output layer and the moment closest to the moment to be predicted. h _RNN is the feature vector extracted by RNN; L is the number of RNN network layers; T is the width of RNN input window.

The multi-dimensional RNN network layer is composed of multiple units in series, and the load state feature information at different times is passed through the hidden state Through transmission, RNN can extract and memorize the dependency features between each time node of the continuous load sequence, and couple the load information of different time nodes. In order to enhance the ability of the RNN network to extract and represent the time-series feature information of the load, this embodiment stacks multiple RNNs to learn the complex nonlinear relationship between the industry load and its influencing factors. The RNN module of the day-ahead power load forecasting based on RNN is shown in Figure 4.

Step 4: Based on the Attention module and RNN module formed in Step 2 and Step 3, build a serial and parallel model integration framework, and use the Attention mechanism and RNN network characteristics to mine the time correlation and long-term dependence of load timing from historical data Features, forming a short-term power load forecasting method based on Attention-RNN;

Based on the Attention module and the RNN module, two model integration frameworks, serial and parallel, are constructed. In the serial integration framework: firstly, input the data in the history window into the RNN module, process it through the RNN module, and extract the time series dependent features of the load sequence; then, extract the feature vector group obtained by the RNN module Input the Attention module, extract the correlation between each time step to obtain the output matrix H considering the correlation of each time step; finally, input H and the synchronous features in the prediction window into the fully connected layer, and output the short-term power load forecast value, serial The schematic diagram of the row integration framework is shown in Figure 5. In the parallel integration framework: first, input the data in the history window into the RNN module and the Attention module in parallel, and mine the time correlation and long-term dependence characteristics of the load sequence; then, extract the feature vector group obtained by the RNN /> The Attention output H and the synchronization features in the prediction window are input to the fully connected layer, and the output is the short-term power load forecast value. The schematic diagram of the parallel integration framework is shown in Figure 6.

In order to evaluate the prediction effect of the model, the root mean square error and the average absolute percentage error are used as the prediction evaluation indicators, as shown in formula (18) and formula (19), respectively.

In order to verify the effectiveness and accuracy of the proposed short-term power load forecasting method based on Attention-RNN, an example simulation is carried out with 96 points of daily load data in a city in 2021.

First, the original 96-point daily load data is preprocessed according to the characteristic representation method of multivariate heterogeneous data proposed in step 1 and the basic framework of single-step forecasting and multi-step forecasting of the day-ahead power load forecasting, including data normalization, feature Encoding, data shaping, and sliding windowing.

Next, based on the sliding window data, the RNN module and the Attention module are formed according to the RNN feature extraction method and the Attention timing correlation quantification method mentioned in Step 2 and Step 3.

Then, according to the two frameworks proposed in step 3, integrate the RNN module and the Attention module, while considering the influence of synchronization-related features in the prediction window, and input it into the fully connected layer. In order to improve the training efficiency of the model while ensuring a certain mapping ability of the model, the number of hidden layers in the GRU module is set to 2, the number of Attention layers is 1, the number of synchronous feature nonlinear mapping layers is 2, and the number of fully connected layers is 1. The modeling work of the daily load forecasting model is completed. On the basis of the determination of the model structure, the data set after sliding window processing is divided into a training set and a test set. The training set is used to train the constructed model and determine the model parameters; verify.

In this embodiment, the Attention module and the RNN module are combined with the serial and parallel model integration framework, and the Attention mechanism and RNN network characteristics are used to consider the load timing characteristics and external multi-dimensional influencing factors. Compared with the traditional prediction method, it has the following advantages:

Consider external multi-dimensional influencing factors: Traditional forecasting methods often only consider historical load data itself, but this technical solution can effectively consider external multi-dimensional influencing factors, such as temperature, humidity, holidays, etc., by introducing the Attention mechanism, which improves the forecasting accuracy.

Fully mine the time series features: By using the RNN network, the time series features of the load data can be modeled, so as to better capture the periodicity, trend and other information in the load data, and further improve the prediction accuracy.

Model fusion improves robustness: Through the serial and parallel model integration framework, the Attention module and the RNN module are fused to fully tap the complementarity between different models and improve the robustness and stability of predictions.

Strong interpretability: Due to the adoption of the Attention mechanism, the importance of each external influencing factor in the prediction can be displayed visually, thereby improving the interpretability of the model.

In order to verify the effectiveness and superiority of the proposed method of the present invention, construct other typical prediction models, such as ANN, CNN, GRU, CNN-GRU, etc., and compare with the prediction results of the present invention, and the evaluation indexes of each prediction model are compared in the following table shown.

Table 1 Comparison of short-term load forecasting model evaluation indicators

The short-term load forecasting curve based on the above model is shown in Fig. 7. According to Figure 7 and Table 1, compared with other short-term load forecasting models, the short-term load forecasting model based on Attention-RNN proposed by the present invention has relatively higher forecasting accuracy, and the Attention-RNN short-term load forecasting model accuracy of the serial architecture is The highest, the smallest error, and the best generalization performance of the model.

The short-term power load forecasting method based on Attention-RNN shown above is a specific embodiment of the present invention, which has already reflected the substantive features and progress of the present invention, and can be used according to actual needs under the inspiration of the present invention. The equivalent modification of its shape, structure and other aspects are all within the scope of protection of this scheme.

Claims (5)

1. A short-term power load forecasting method based on Attention-RNN, is characterized in that, comprises the following steps: 1) Determine the characteristic representation method of multivariate heterogeneous data including temperature and holiday information, define the data set format for short-term load forecasting considering multiple load influencing factors, and construct the basic framework of single-step forecasting and multi-step forecasting for day-ahead power load forecasting ; 2) According to the basic framework of day-ahead power load forecasting formed in step 1), based on the Attention mechanism, quantify the time-series correlation implied between each time node in the load time series, and form the day-ahead power load forecasting Attention module; 3) According to the basic framework of the day-ahead power load forecasting formed in step 1), based on the RNN, the hidden trend features and cycle features in the long-term load sequence are extracted, and the timing dependence of the load sequence is mined to form the RNN module of the day-ahead power load forecasting; 4) Based on the Attention module and RNN module formed in step 2) and step 3), build a serial and parallel model integration framework, and use the Attention mechanism and RNN network characteristics to mine the time correlation and long-term load timing from historical data. Depending on the feature, a short-term power load forecasting model based on Attention-RNN is formed, and short-term power load forecasting is performed according to the short-term power load forecasting model. 2. a kind of short-term power load forecasting method based on Attention-RNN according to claim 1, is characterized in that: in step 1), original data set is expressed as: In the formula: X is the input matrix; x _i is the input feature vector of the i-th day, which is composed of the 96-point load feature vector l _i and the external factor feature vector f _i ; l _i ∈ R ⁹⁶ is the load vector of the i-th day, the number of points Depends on the sampling frequency; f _i ∈ R ²⁶ is the external factor feature vector of the i-th day, coded by weekday type d _i ∈ R ⁷ , season type code s _i ∈ R ⁴ , monthly type code m _i ∈ R ¹² , holiday The type code h _i ∈ R ² and the meteorological feature vector n _i are composed, and f _i is expressed as: f _i ＝[d _i s _i m _i h _i n _i ] (2) Among them, the time-scale feature encoding method adopts one-hot encoding; after determining the data structure of short-term power load forecasting, the basic framework of single-step forecasting and multi-step forecasting of day-ahead power load forecasting is expressed as: X _i+1:i+T ＝[x _i+1 x _i+2 ... x _i+T ] ^T (5) F _{i+T+K:i+T+K+τ} ＝[f _i+T+K f _i+T+K+1 ... f _i+T+K+τ ] ^T (6) In the formula, is the power load forecasting matrix; X _i+1:i+T is the historical input matrix; F _{i+T+K:i+T+K+τ} is the synchronous feature matrix; T and τ are the historical window width and forecast window width ; K is the number of forecast days in advance; Φ is the short-term power load forecasting model; θ ^* is the parameters obtained from model training. 3. A short-term power load forecasting method based on Attention-RNN according to claim 1, characterized in that: in step 2), the implicit temporal correlation between each time node in the load time series is quantified based on the Attention mechanism to extract cross-correlation features; in order to quantify the time correlation between each time step of the load time series, the "key-value" mechanism is adopted; the input feature vector is transformed into query vector q _i , key vector k _i and value vector v _i through linear mapping ; The correlation score a between different time steps is obtained by calculating the dot product of the query vector and the key vector: a=q·k (7) The time correlation between different time steps is obtained by calculating the dot product of the query vector and the key vector at different time steps. The overall process of quantifying the time series correlation of loads based on the Attention mechanism is expressed as: q _i =W ^q x _i , _ki =W ^k x _i ,v _i =W ^v x _i ,a _i,j =q _i k _j (8) A＝softmax(K ^T Q), H＝VA (10) In the formula, W ^q , W ^k , W ^v are the query mapping matrix, key mapping matrix and value mapping matrix; a _{i, j} are the correlation scores between the input vectors x _i and x _j ; A matrix consists of a _{i, j} Composition; K and Q are the key matrix and the query matrix, which are composed of ki _and q _i respectively; H is the Attention output. 4. A short-term power load forecasting method based on Attention-RNN according to claim 1, characterized in that: in step 3), the processing formula of the day-ahead power load forecasting RNN module is formed based on the load time series dependence characteristics of RNN As shown in (11)-(14); In the formula, is the weight matrix of the RNN layer l;/> is the bias of the first layer of RNN; σ is the sigmoid activation function; W _o is the linear mapping matrix; /> The hidden state features at time t extracted for the l-layer RNN network; /> is the hidden state feature of the RNN output layer and the moment to be predicted; h _RNN is the feature vector extracted by RNN; L is the number of RNN network layers; T is the width of the RNN input window. 5. a kind of short-term power load forecasting method based on Attention-RNN according to claim 1, is characterized in that: in step 4): original data set generates history window and prediction window by sliding window processing; In the serial integration framework: firstly, input the data in the history window into the RNN module, process it through the RNN module, and extract the time series dependent features of the load sequence; then, extract the feature vector group obtained by the RNN module Input the Attention module to extract the correlation between each time step. Finally, input H and the synchronous features in the prediction window into the fully connected layer, and output the short-term power load prediction value; In the parallel integration framework: first, input the data in the history window into the RNN module and the Attention module in parallel, and mine the time correlation and long-term dependence characteristics of the load sequence; then, the feature vector group extracted by the RNN The Attention output H and the synchronous features in the prediction window are input into the fully connected layer, and the output is the short-term power load forecast value.