Short-term power load prediction method considering multiple factors

Abstract

The invention discloses a short-term power load prediction method considering multiple factors, and the method comprises the following steps: 1, obtaining historical power load data and external factor data, carrying out the correlation analysis, obtaining the power load data and external factor data with the correlation meeting the requirements, and constructing a training set; step 2, constructing a DCR-Attention-BiLSTM combined model, wherein the DCR-Attention-BiLSTM combined model comprises an expansion convolution residual block DCR, a time attention mechanism, a cross attention mechanism and a BiLSTM network; using the training set in the step 1 to train the combined model; and during training, outputting a short-term power load data prediction result by the combined model. According to the method, the influence of irrelevant external factors on the load prediction precision can be reduced, the prediction precision of the power load can be effectively improved, and the problem of high difficulty in model parameter adjustment is avoided.

Description

Technical Field

[0001] The invention relates to the field of power load forecasting methods, in particular to a short-term power load forecasting method taking multiple factors into consideration. Background Art

[0002] Electricity load forecasting plays an important role in the new power system. Accurate load forecasting helps improve energy utilization, reduce unnecessary energy waste, and has certain economic benefits. According to the length of the forecast time, electricity load forecasting can be divided into three types: short-term load forecasting, medium-term load forecasting, and long-term load forecasting. Among them, short-term electricity load forecasting is a scientific and accurate forecast of electricity demand in the next few hours to days, which can ensure the balance between electricity supply and demand, reduce the waste of electricity resources, and improve the stability and safety of the power system.

[0003] In the existing technology, some hybrid prediction models are used to predict short-term power loads. For example, in the literature [LU Jixiang, ZHANG Qipei, YANG Zhihong, et al. Short-term Load Forecasting Method Based on CNN-LSTM Hybrid Neural Network Model [J]. Automation of Electric Power Systems, 2019, 43 (8): 131-137. LU Jixiang, ZHANG Qipei, YANG Zhihong, et al. Short-term Load Forecasting Method Based on CNN-LSTM Hybrid Neural Network Model [J]. Automation of Electric Power Systems, 2019, 43 (8): 131-137. (in Chinese)], CNN and LSTM are combined, CNN is first used to complete the feature extraction between non-continuous data, and then LSTM is used for short-term load prediction, which achieves good prediction results.

[0004] In the literature [SEKHAR C, DAHIYA R. Robust framework based on hybrid deep learning approach for short term load forecasting of building electricity demand [J]. Energy, 2023, 268: 126660.], a residential electricity consumption prediction model based on a hybrid of CNN and BiLSTM was proposed. The gray wolf optimization algorithm was used to select the optimal parameter set, which achieved better prediction accuracy compared with the single BiLSTM model and the combined CNN-LSTM model.

[0005] However, although the hybrid prediction model of the prior art can improve the prediction accuracy by using multiple CNNs, the excessive number of network layers will make it more difficult to adjust the parameters of the prediction model. In addition, when faced with load data and its multiple external factors, the above-mentioned model directly extracts feature information from the load data and its external factors as a whole, resulting in the newly acquired feature information including the relationship between external factors, introducing additional feature information and resulting in information redundancy, which limits the accuracy of power load prediction to a certain extent. Summary of the invention

[0006] The present invention provides a short-term power load forecasting method taking multiple factors into consideration, so as to solve the problems in the prior art of using a hybrid forecasting model for forecasting, that is, the model parameter adjustment is difficult and the redundant information affects the accuracy of the forecasting result.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] The short-term power load forecasting method considering multiple factors includes the following steps:

[0009] Step 1: Obtain historical power load data and external factor data that affect the historical power load data, perform correlation analysis on the power load data and the external factor data, obtain power load data and external factor data whose correlation meets the requirements, and construct a training set with the power load data and external factor data whose correlation meets the requirements;

[0010] Step 2: construct a DCR-Attention-BiLSTM combination model, wherein the DCR-Attention-BiLSTM combination model includes a dilated convolution residual block DCR, a temporal attention mechanism, a cross attention mechanism, and a BiLSTM network;

[0011] The DCR-Attention-BiLSTM combined model is trained using the training set in step 1. During training, the dilated convolution residual block DCR extracts power load features from the power load data in the training set and extracts external factor features from the external factor data in the training set. The cross attention mechanism extracts correlation features based on the power load features and external factor features. The time attention mechanism extracts time features based on the power load features. The time features and correlation features are input into the BiLSTM network. The BiLSTM network further captures the bidirectional time features of the input data. Finally, the linear layer directly outputs the short-term power load data prediction results.

[0012] In a further step 1, the Pearson correlation coefficient method is used to perform correlation analysis on the power load data and the external factor data.

[0013] In a further step 1, the power load data and external factor data whose correlation meets the requirements are normalized to construct a data set, and the data in the data set are divided into a test set and the training set, wherein the training set is used for training in step 2, and the test set is used to evaluate the performance of the DCR-Attention-BiLSTM combination model after training in step 2.

[0014] In the further step 2, the dilated convolution residual block DCR includes a dilated convolutional network and a residual connection block. After the dilated convolutional network extracts features from the input data, the residual connection block activates the extracted features and then performs residual connection, thereby obtaining the final output features.

[0015] In the further step 2, in the BiLSTM network, the time feature and the correlation feature are first subjected to BiLSTM feature extraction respectively, and then the features obtained after the BiLSTM feature extraction are concatenated and then subjected to BiLSTM feature extraction again, and finally the linear layer outputs the short-term power load data prediction result.

[0016] The present invention combines the dilated convolutional network (DCN), the residual-refinement (Res) method, the attention mechanism and BiLSTM, and proposes a short-term power load forecasting method combining DCR-BiLSTM and the cross and temporal attention mechanism.

[0017] In the present invention, the Pearson correlation coefficient (PCCs) is firstly used to measure the dependency between external factor data including temperature, humidity and electricity price and power load data, and external factors with strong correlation are selected as input data to reduce the influence of irrelevant external factors; secondly, the local pattern features of power load data and external factor data are captured respectively by dilated convolution residual block DCN, and multiple feature extractions are performed using residual structure to mine deeper input data rules; then, the time attention mechanism is introduced to deeply mine the time features in historical load data, and the cross attention mechanism is introduced to deeply mine the correlation features of power load features and external factor features, so as to more comprehensively capture the nonlinear relationship between external factors and power load data; finally, the extracted features are used for load forecasting through BiLSTM network, and the bidirectional time features of input data are further captured through BiLSTM, and finally the short-term power load data forecasting result is directly output by the linear layer to realize accurate prediction of load data.

[0018] Compared with the prior art, the present invention uses the Pearson correlation coefficient to analyze the correlation between power load data and external factors, which can reduce the impact of irrelevant external factors on load prediction accuracy. The present invention uses the time attention mechanism and the cross self-attention mechanism to extract the time characteristics in the load data and the nonlinear correlation characteristics between the external factor characteristics and the load data, which can effectively improve the prediction accuracy of power load. In addition, since the present invention does not use CNN, it avoids the problem of difficulty in adjusting model parameters.

[0019] The present invention is verified by numerical simulation, and the experimental results show that the method proposed in the present invention has lower error and higher prediction accuracy compared with other single prediction and combined prediction comparison models. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of a method according to an embodiment of the present invention.

[0021] Figure 2 It is a diagram of the DCR-Attention-BiLSTM combined model architecture in an embodiment of the present invention.

[0022] Figure 3 It is a BiLSTM architecture diagram in an embodiment of the present invention.

[0023] Figure 4 It is a heat map of the Pearson correlation coefficient matrix between the external factors and the power load in the experiment.

[0024] Figure 5 This is a comparison chart of the prediction results of different models in the experiment, where: (a) is the result of one week; (b) is the result of one day. DETAILED DESCRIPTION

[0025] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0026] like Figure 1 As shown, this embodiment discloses a short-term power load forecasting method considering multiple factors, comprising the following steps:

[0027] Step 1: Obtain historical power load data and external factor data that affect the historical power load data, perform correlation analysis on the power load data and the external factor data, and obtain power load data and external factor data whose correlation meets the requirements.

[0028] In addition to the influence of the load data itself, the power load data is also affected by a variety of other external factors, such as weather factors, economic factors, major holidays, etc., which makes the power load data present strong randomness and volatility. Therefore, in this embodiment, in order to explore the strong correlation characteristic relationship between external factors and load target variables, this embodiment adopts the Pearson correlation coefficient method to analyze the correlation between external factor data and power load data. The analysis formula is shown as follows:

[0029]

[0030] Where: X i is the value of the ith external factor; Y i is the ith power load data value; and They are X i and Y i The mean of ; n is the sample size of the selected data; r(X,Y) is the correlation coefficient, and its value range is [-1,1]. The larger the absolute value of the correlation coefficient, the stronger the linear correlation between the two variables.

[0031] Therefore, in this embodiment, the Pearson correlation coefficient of the power load data and the external factor data is calculated, and the external factor data whose Pearson correlation coefficient meets the requirements is selected, that is, the external factor data and power load data whose correlation meets the requirements are selected to construct a data set for subsequent use.

[0032] After obtaining the data set, the data in the data set is normalized, and then the data in the data set is divided into training set data and test set data according to the set ratio, thereby obtaining the training set and the test set.

[0033] Step 2: Build a DCR-Attention-BiLSTM combined model. The model architecture is as follows: Figure 2 As shown in the figure, the DCR-Attention-BiLSTM combination model includes the dilated convolution residual block DCR, the temporal attention mechanism, the cross attention mechanism, and the BiLSTM network. The DCR-Attention-BiLSTM combination model is described as follows:

[0034] (A1) Dilated Convolution Residual Block DCR

[0035] In order to fully extract historical data information and long-distance dependencies in time data and enhance the model's ability to extract information from complex time data, the dilated convolution residual block DCR is introduced in this embodiment. The dilated convolution residual block DCR includes a dilated convolution network and a residual connection block to better retain historical time data information.

[0036] The dilated convolutional network is a technology that expands the receptive field based on the convolutional neural network. By introducing the concept of "expansion rate", the sampling points inside the convolution kernel are no longer continuous, but are sampled at a certain time interval, so that a larger receptive field can be achieved through a small number of convolutional layers, which can effectively extract local nonlinear features between time series without increasing the computational cost. The calculation process of the dilated convolutional network is:

[0037]

[0038] Where: K is the size of the convolution kernel; w i is the i-th weight of the convolution kernel; s (t+i·d) is the value at time t+i·d in the input time series s; d is the dilation rate, which indicates the spacing between convolution kernel elements. S is the output matrix after feature extraction.

[0039] As the depth of the network layer increases, the prediction model may have a series of potential problems such as gradient disappearance and gradient explosion, which will cause information loss in the prediction model and network instability. The residual connection can solve the above problems well. Therefore, under the premise of equal number of network layers, this embodiment introduces a residual structure to make the model converge faster. The calculation process of this module is shown in the following formula:

[0040] S1=ReLu(S)

[0041] S2=S1+cov

[0042] Where: ReLu(·) is the activation function; S1 is the data after activation; cov is the residual connection data; S2 is the output matrix of the residual block, that is, the corresponding features of the output. It should be noted that if the shape of cov data is the same as that of S1, the two are directly added; if the shapes are different, regular convolution CNN is used to change its shape to meet the requirements of residual connection.

[0043] (A2) Cross-Attention Mechanism

[0044] When predicting power load data, the more external factors the input data has, the better the predictive performance of the model will be. However, different external factors have different effects on the power load prediction results. Therefore, this embodiment uses an attention mechanism to dynamically weight and capture the correlation between load data and input features.

[0045] As an important variant of the attention mechanism, the cross-attention model can analyze the correlation between different input feature factors, assign weights to the input load data and external feature variables, and enhance the prediction model's ability to extract the relationship between important external features and load data, thereby fully capturing the impact of different external features on power load data. The specific calculation process of the cross-self-attention score is shown in the following formula:

[0046]

[0047] Where: Q, K, V are the query matrix, key matrix, and value matrix respectively; d k is the dimension of the key; is a scaling factor used to control the gradient of the attention mechanism; softmax(·) is an activation function, indicating normalization; Attention out (Q, K, V) are the correlation feature results output by the cross-attention mechanism.

[0048] (A3) Temporal Attention Mechanism

[0049] Different from the traditional attention mechanism, the temporal attention mechanism can selectively focus on the impact of different time steps on the power load forecast results. It assigns different weights to time steps by using weighted scoring, and pays special attention to time steps with high weights to obtain key time step features. The temporal attention feature can better analyze the time dependency within the power load data and reduce the impact of non-important time features. The calculation process is as follows:

[0050] E = tanh(WQ+UK)

[0051]

[0052] Where: E is the similarity matrix; W and U are learnable matrices; tanh(·) is the activation function; Q and K are the query matrix and key matrix respectively; α i is the attention mechanism score weight of the current position, e i and e j For each element in E; K i represents each eigenvalue in K; N E and N K Respectively represent the number of elements in E and K; Attention time (K) is the temporal feature finally output by the temporal attention mechanism.

[0053] (A4) BiLSTM Network

[0054] In time series data prediction, LSTM can capture the dependencies between time series and effectively solve the problems of gradient explosion and gradient disappearance. However, using LSTM to perform unidirectional load forecasting on time series makes it difficult to capture the dependencies of long time series data, and lacks in prediction accuracy. The bidirectional long short-term memory network BiLSTM processes the input data from the forward and reverse directions respectively, while considering the past and future information of the input data, so that the model can better and more comprehensively capture the long-term dependencies in the input time series data. Therefore, this embodiment uses the bidirectional long short-term memory network BiLSTM for power load forecasting. The internal structure of BiLSTM is as follows: Figure 3 shown.

[0055] Figure 3 In the above example, for each BiLSTM unit, its output at time t is jointly affected by the LSTM output of the forward propagation at time t and the LSTM output of the backward propagation at time t, which are linearly superimposed according to certain calculation weights. The calculation process of each BiLSTM unit is shown in the following formula:

[0056]

[0057] Where: x t is the input at the current moment; and Represent the output of the hidden layer at time t and the forward propagation and reverse propagation respectively; and They represent the hidden layer state at time t-1 of the forward LSTM propagation and the hidden state at the t+1th input time step in the reverse LSTM layer; LSTM(·) represents the calculation process of LSTM; W1 is the hidden layer output weight matrix of the forward propagation; W2 is the hidden layer output weight matrix of the reverse propagation; b is the hidden layer weight optimization parameter at time t; H t Represents the output of the hidden layer of the BiLSTM model at time t.

[0058] (A5) Load forecasting result output

[0059] A linear layer is connected after the BiLSTM layer to directly perform a linear transformation on the output result of the upper layer, converting the output load forecast result into one-dimensional data to obtain the final forecast value Y.

[0060] Y=WB+b B

[0061] Where: B is the output of BiLSTM; W is the weight matrix; b B Optimize parameters for weights.

[0062] In this embodiment, after constructing the DCR-Attention-BiLSTM combination model, the DCR-Attention-BiLSTM combination model is trained using the training set in step 1. During training, the data processing process of the DCR-Attention-BiLSTM combination model is as follows:

[0063] (B1) First, the dilated convolution residual block DCR performs preliminary feature extraction on the power load data and external factor data in the training set, extracts power load features from the power load data, and extracts external factor features from the external factor data in the training set, so as to retain more of the history of the input sequence and reduce the impact of missing data information on the prediction model. The feature extraction process of the dilated convolution residual block DCR in this embodiment is shown in the following formula:

[0064] X1=F DCR (X)

[0065] Y1=F DCR (Y)

[0066] Where X1 and Y1 are the external factor characteristics and power load characteristics of the dilated convolution residual block DCR output respectively; F DCR (·) represents the feature extraction process of DCR; X represents the input external factor data; Y represents the input short-term power load data.

[0067] (B2) The initial data feature extraction of the dilated convolution residual block DCR is mainly to effectively retain the historical time information of the load data, but it does not fully capture the relationship between the time step and the load. Therefore, this embodiment adopts a time attention mechanism to further extract the time information of the power load characteristics obtained by the dilated convolution residual block DCR, aiming to selectively focus on the information of important time steps, ignore the impact of irrelevant time steps on load data prediction, and obtain time features. The feature extraction process of the time attention mechanism in this embodiment is shown as follows:

[0068] Y time =Attention time (Y1)

[0069] Where: Y time Temporal features output by the temporal attention mechanism.

[0070] (B3) Using the cross attention mechanism, the correlation between the power load feature obtained by the dilated convolution residual block DCR and the external factor feature is extracted, thereby capturing the relationship between the power load and different external factors, so that the correlation between the external factors and the power load is fully extracted to obtain the correlation feature. The feature extraction process of the cross attention mechanism in this embodiment is shown in the following formula:

[0071] Y cor =Attention out (Y1,X1,X1)

[0072] Where: Y cor Correlation features output by the cross-attention mechanism.

[0073] (B4) Finally, the BiLSTM network is used to further extract the time features and correlation features, so that the DCR-Attention-BiLSTM combination model can fully capture the bidirectional time features of the input sequence, learn deeper time series features, and the prediction results are directly output by the linear layer. The feature extraction process of the BiLSTM network in this embodiment is shown in the following formula:

[0074] Y2=F BiLSTM (Y time )

[0075] Y3=F BiLSTM (Y cor )

[0076] Y cat =Concat(Y2,Y3)

[0077] Y4=F BiLSTM (Y cat )

[0078] Y out =FC(Y4+b)

[0079] Where: F BiLSTM (·) represents the BiLSTM extraction process; Y2, Y3, and Y4 represent the outputs after BiLSTM feature extraction; Concat(·) represents the concatenation operation; Y cat represents the output after concatenation; FC(·) is the linear output function; Y out This is the final power load data prediction result of the BiLSTM network.

[0080] In the training process of this embodiment, the MSE loss function is selected to measure the difference between the predicted value and the true value, and the parameters of the DCR-Attention-BiLSTM combination model are updated based on the change in the size of the loss function, thereby obtaining the trained DCR-Attention-BiLSTM combination model.

[0081] After training in this embodiment, the test set is used to evaluate the performance of the DCR-Attention-BiLSTM combination model after training, and the difference between the predicted power load data and the actual power load data is measured by the set evaluation index, and the model performance is evaluated based on this.

[0082] The performance of the method of this embodiment is further explained below in combination with numerical simulation experiments.

[0083] This experiment will verify the effectiveness of the DCR-Attention-BiLSTM combination model proposed in the above embodiment on the Python simulation platform. First, the parameter settings, data sets and evaluation index selection are introduced, and then the selection of external influencing factors and the effectiveness verification of the proposed prediction model are given, and the performance is compared with the single prediction model and the hybrid prediction model.

[0084] (1) Parameter settings and data sets

[0085] The DCR-Attention-BiLSTM combined model and its comparison method mentioned in the embodiment are completed in an experimental environment where the hardware device is NVIDIA GeForce RTX 4090 and the deep learning framework is Python 3.8, Pytorch1.8 and Anaconda3.

[0086] This experiment uses the Australian electricity load and price forecasting dataset as a sample dataset. The dataset is collected every 30 minutes, from January 1, 2006 to December 31, 2010, with a total of 87,646 data samples. In addition to the real power load data, the dataset also includes five external factors: dry bulb temperature, wet bulb temperature, dew point temperature, humidity, and electricity price. This experiment selects 70% of the dataset as a training set to train the prediction model, 15% as a validation set for model tuning, and 15% as a test set to verify the final prediction effect.

[0087] At the same time, in order to ensure the accuracy of the prediction results of the model, this experiment achieves the optimal hyperparameter selection by setting different hyperparameter combinations. After multiple model pre-training and hyperparameter optimization adjustments, the hyperparameter selection settings of the DCR-Attention-BiLSTM combination model proposed in the embodiment are shown in Table 1.

[0088] Table 1 Hyperparameter settings

[0089]

[0090] (2) Evaluation indicators

[0091] In order to evaluate the prediction effect of the model proposed in the embodiment, this experiment selects the mean absolute percentage error (MAPE), root mean square error (RMSE) and mean absolute error (MAE) as the evaluation indicators of the prediction effect of the model. The smaller the values ​​of MAPE, RMSE and MAE, the smaller the prediction error of the model and the higher the prediction accuracy. The specific indicator definition is shown in the following formula:

[0092]

[0093] Where: y i represents the true value of the i-th sample data; N is the total number of samples in the power load data set; y i ' represents the predicted value of the i-th sample data.

[0094] (3) Selection of influencing factors

[0095] In the data set selected for this experiment, the characteristic parameters that affect the load data are four weather factors: dry bulb temperature, wet bulb temperature, dew point temperature, and humidity, and one economic factor: electricity price. The analysis results based on the Pearson correlation coefficient are as follows: Figure 4 As shown in the figure, the redder the color, the higher the correlation. Figure 4 It can be seen that the absolute values ​​of the correlation coefficients between dry-bulb temperature (Dry-bulb T), dew point temperature (Dew point T), humidity (Humidity), electricity price (Electricity price) and power load data are all greater than 0.1, which have a strong correlation; while the absolute value of the correlation between warm-bulb humidity (Wet-bulb T) and power load data is 0.02, which is relatively the lowest correlation. Therefore, this experiment selects the other four external influencing factors and power load data except warm-bulb humidity as the input features of the prediction model.

[0096] (4) Prediction results and analysis

[0097] 4.1) Comparison with a single prediction model

[0098] In order to verify the effectiveness of the DCR-Attention-BiLSTM combined model proposed in the embodiment in short-term power load forecasting considering multiple factors, this experiment compares it with the existing single prediction models TCN, BiGRU and BiLSTM models widely used in power load forecasting. In this experiment, in order to ensure the accuracy and rationality of the comparative experiment, the model performance comparison is carried out under the same data set and configuration parameters, and the comparison results are shown in Table 2.

[0099] Table 2 Comparison of prediction results of single model

[0100]

[0101] From the analysis of the prediction results in Table 2, it can be seen that the DCR-Attention-BiLSTM combination model proposed in the embodiment has a MAPE reduction of 43.40%, 24.13% and 11.94% respectively, RMSE reduction of 52.53%, 12.85% and 4.83% and MAE reduction of 56.35%, 13.47% and 6.72% in the three indicators evaluated compared with the comparison models TCN, BiGRU and BiLSTM. The experimental results show that the DCR-Attention-BiLSTM combination model proposed in the embodiment has better evaluation indicators, and the proposed prediction model has higher prediction accuracy than the single prediction model.

[0102] 4.2) Comparison with other combination models

[0103] In order to further illustrate the effectiveness of the DCR-Attention-BiLSTM combination model proposed in the embodiment in short-term power load forecasting, this experiment compares the DCR-Attention-BiLSTM combination model proposed in the embodiment with the combination models CNN-LSTM, CNN-BiLSTM, CNN-BiLSTM-Attention and BiLSTM-LSTM that are currently commonly used for short-term power load forecasting. The prediction results are shown in Table 3.

[0104] Table 3 Comparison of prediction results of combined models

[0105]

[0106]

[0107] From the analysis of the results in Table 3, it can be seen that the evaluation index of the DCR-Attention-BiLSTM combined model proposed in the embodiment is compared with the CNN-LSTM, CNN-BiLSTM, CNN-BiLSTM-Attention, and BiLSTM-LSTM methods. The MAPE is reduced by 34.57%, 14.04%, 11.75%, and 14.72%, respectively, the RMSE is reduced by 35.79%, 10.63%, 10.06%, and 12.38%, respectively, and the MAE is reduced by 38.00%, 11.38%, 10.56%, and 13.18%, respectively. After analysis, it is known that the DCR-Attention-BiLSTM combined model proposed in the embodiment has the smallest prediction error among the compared combined prediction models, which further proves the effectiveness of the proposed model in short-term power load forecasting.

[0108] 4.3) Visualization of results

[0109] In order to more intuitively display the prediction results of power load, this experiment selected 336 data points from the last week and the last day of the test data set for the results display. Figure 5 shown.

[0110] Depend on Figure 5 (a) It can be seen that compared with other comparison models, the curve fitting degree of the DCR-Attention-BiLSTM combination model proposed in the embodiment is closer to the actual power load value, and it can better fit the actual power load data at both the gentle part of the load change and the peak and valley values. Figure 5 In (b), the changes in the peak and trough of the power load within a day can be more intuitively observed. Compared with the combined prediction model, the single prediction model has a large deviation between the peak and trough prediction values ​​and the true values, and the prediction performance is poor. The DCR-Attention-BiLSTM combined model proposed in the embodiment has the best prediction effect.

[0111] The experimental results prove that the DCR-Attention-BiLSTM combined model proposed in the embodiment can accurately learn the internal change rules of power load data and the impact of external variables on load data, and has better prediction performance.

[0112] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0113] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A short-term power load forecasting method considering multiple factors, characterized in that: The following steps are involved: Step 1: Obtain historical power load data and external factor data that affect the historical power load data, perform correlation analysis on the power load data and the external factor data, obtain power load data and external factor data whose correlation meets the requirements, and construct a training set with the power load data and external factor data whose correlation meets the requirements; Step 2: construct a DCR-Attention-BiLSTM combination model, wherein the DCR-Attention-BiLSTM combination model includes a dilated convolution residual block DCR, a temporal attention mechanism, a cross attention mechanism, and a BiLSTM network; The DCR-Attention-BiLSTM combined model is trained using the training set in step 1. During training, the dilated convolution residual block DCR extracts power load features from the power load data in the training set and external factor features from the external factor data in the training set. The cross attention mechanism extracts correlation features based on the power load features and external factor features. The time attention mechanism extracts time features based on the power load features. Finally, the time features and correlation features are input into the BiLSTM network. The BiLSTM network further captures the bidirectional time features of the input data. Finally, the linear layer directly outputs the short-term power load data prediction results.

2. The short-term power load forecasting method considering multiple factors according to claim 1 is characterized in that: In step 1, the Pearson correlation coefficient method is used to perform correlation analysis on the power load data and external factor data.

3. The short-term power load forecasting method considering multiple factors according to claim 1 is characterized in that: In step 1, the power load data and external factor data whose correlation meets the requirements are normalized to construct a data set, and the data in the data set are divided into a test set and the training set, wherein the training set is used for training in step 2, and the test set is used to evaluate the performance of the DCR-Attention-BiLSTM combination model after training in step 2.

4. The short-term power load forecasting method considering multiple factors according to claim 1 is characterized in that: In step 2, the dilated convolution residual block DCR includes a dilated convolutional network and a residual connection block. After the dilated convolutional network extracts features from the input data, the residual connection block activates the extracted features and then performs residual connection, thereby obtaining the final output features.

5. The short-term power load forecasting method considering multiple factors according to claim 1 is characterized in that: In step 2, in the BiLSTM network, BiLSTM feature extraction is first performed on the time feature and the correlation feature, respectively, and then the features obtained after the BiLSTM feature extraction are concatenated and BiLSTM feature extraction is performed again, and finally the linear layer outputs the short-term power load data prediction result.

Images