Bidirectional long short-term memory network-based power load prediction method and apparatus

Abstract

The present invention belongs to the field of power load prediction. Disclosed are a bidirectional long short-term memory network-based power load prediction method and apparatus. The method comprises: first, acquiring power data, preprocessing same, and dividing same into a training set and a test set; second, creating a preliminary feature extraction unit and extracting preliminary features; then, combining the preliminary features and an input, and inputting same into a prediction unit to acquire a prediction result; and finally, carrying out training by means of the training set, and using the test set for testing the effect of a power load prediction model. The apparatus comprises a data module, a prediction module and an output module. The present invention can efficiently and accurately predict power loads.

Description

Power load forecasting method and device based on bidirectional long short-term memory network Technical Field

[0001] The present invention belongs to the field of power load forecasting, and in particular relates to a method and device for power system load forecasting based on a bidirectional long short-term memory network. Background Art

[0002] The electricity generated by the power system is difficult to store, so its production and consumption are essentially synchronized. However, electricity consumption is not always stable. For various reasons, the power load often fluctuates significantly in a short period of time, making it difficult for the grid system to make appropriate scheduling arrangements, resulting in insufficient power production and supply or energy waste. Therefore, effective power load forecasting is necessary.

[0003] At present, with the further development of deep learning technology, the continuous expansion of the scale of power systems and the continuous improvement of information technology, power grid data collection has become more and more accurate, providing reliable and massive data for load forecasting, which is conducive to the application of deep learning in power loads.

[0004] Furthermore, traditional mathematical methods are limited to linear prediction, making them difficult to predict for nonlinear and highly volatile data such as power systems. Consequently, research on power load forecasting using convolutional networks or long-short-term memory networks is growing. However, these methods are ineffective for periodic power forecasting with limited sample data points, suffer from low accuracy, are significantly affected by interference, cannot respond promptly to power load changes, and exhibit poor generalization.

[0005] Summary of the Invention

[0006] In response to the above problems, the present invention proposes a power load forecasting method and device based on a bidirectional long short-term memory network. Preliminary feature extraction is performed through two feature extractors composed of a deep separable convolutional network, and then residual connection is introduced. After splicing, the bidirectional long short-term memory network BiLSTM is used to further perform deep feature extraction and power load forecasting. On the one hand, the number of parameters of the entire network is small, which accelerates the prediction efficiency. On the other hand, accurate prediction of power load is achieved.

[0007] In one aspect, the present invention provides a method for power load forecasting based on a bidirectional long short-term memory network, comprising the following steps:

[0008] Step 1: Obtain power data, preprocess it, and divide it into training and test sets.

[0009] Step 2: Construct a preliminary feature extraction unit to extract preliminary features.

[0010] Step 3: Through residual connection, the preliminary features are concatenated with the input, input into the prediction unit, and the prediction results are obtained.

[0011] Step 4: Train the model using the training set and use the test set to test the power load prediction effect.

[0012] Furthermore, the pre-processing operation in step 1 specifically includes the following processes:

[0013] First, the acquired power data is normalized and scaled to the same range.

[0014] Secondly, because deep learning is extremely dependent on data, eliminating outliers in power data can increase the convergence speed of the neural network and make more accurate predictions.

[0015] Finally, because the data points collected by the power system are limited, it is not conducive to the prediction analysis and extraction of implicit features by neural networks. Therefore, the power data is decomposed into signals through signal decomposition algorithms, and the sample data is expanded to make the prediction more accurate.

[0016] Furthermore, the preliminary feature extraction unit in step 2 is specifically:

[0017] The preliminary feature extraction unit is divided into two paths. Each path contains a feature extractor composed of several depthwise separable convolutions, which performs shallow feature extraction. The two feature extractors have the same structure. The features extracted by the two feature extractors are then element-by-element added to obtain preliminary features. Due to the characteristics of depthwise separable convolution, the number of parameters is significantly reduced compared to conventional convolution, so with fewer parameters, feature extraction performance comparable to conventional convolution can be achieved.

[0018] Furthermore, the prediction unit in step 3 is specifically:

[0019] The prediction unit comprises an n-layer BiLSTM (bidirectional long short-term memory) network, which extracts features at the time series level. The output of the final BiLSTM layer serves as the input to a fully connected network, passing through an m-layer multilayer perceptron (MLP) to output the power load forecast. Due to the BiLSTM's high efficiency in extracting contextual information, the accuracy of power load forecasts has been significantly improved.

[0020] On the other hand, the present invention provides a power load forecasting device based on a bidirectional long short-term memory network, which includes a data module, a forecasting module and an output module.

[0021] The data module is used to obtain power data and perform preprocessing.

[0022] The prediction module includes a preliminary feature extraction submodule and a prediction submodule, which is used to input the data preprocessed by the data module as input, perform power load prediction, and output power load prediction results.

[0023] The preliminary feature extraction submodule extracts preliminary features from the preprocessed power data through two identical feature extractors;

[0024] The input of the prediction submodule is composed of the input of the preliminary feature extraction submodule and the preliminary features through a residual connection, and then the subsequent feature extraction is performed through BiLSTM and multi-layer perceptron, and the power load prediction result is output.

[0025] The output module is used to display the power load forecast result on the display screen to complete human-computer interaction.

[0026] The beneficial effects of the present invention are:

[0027] The present invention performs signal decomposition on the acquired power data, and refines and decomposes the original data while retaining the original information, so as to make the network prediction result more accurate.

[0028] The present invention uses two feature extractors composed of depthwise separable convolutions to perform preliminary feature extraction without significantly increasing parameters, thereby improving prediction efficiency while ensuring improved accuracy. The input samples are then introduced into the BiLSTM model through residual connections, avoiding information loss caused by network deepening, thereby accurately predicting power load. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] FIG1 is a flow chart of a method for power load forecasting based on a bidirectional long short-term memory network according to an embodiment of the present invention;

[0030] FIG2 is a schematic diagram of a network model structure of a power load forecasting method based on a bidirectional long short-term memory network according to an embodiment of the present invention;

[0031] FIG3 is a diagram showing the prediction results of an experimental method for power load forecasting based on a bidirectional long short-term memory network according to an embodiment of the present invention;

[0032] FIG4 is a schematic diagram of a power load forecasting device based on a bidirectional long short-term memory network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0034] On the one hand, the present invention provides a method for power load forecasting based on a bidirectional long short-term memory network. The specific process is shown in Figures 1 and 2. The present invention includes the following steps:

[0035] Step 1: Obtain electricity data within one year, preprocess it, and divide it into training and test sets.

[0036] Through the Australian Energy Market website, we obtained one year of data samples from four states in Australia. One data item was collected every half hour, and the total number of sample data for each state was 17,520, with 4,320, 4,358, 4,416, and 4,416 items in the four quarters respectively.

[0037] First, the power data is normalized to the maximum and minimum values, and the data is scaled to the same range to make network training easier to converge. The formula is as follows:

[0038] Among them, x is the original data, x min is the minimum value of the data point set, x max is the maximum value of the data point set, x′ is the normalized value, and the value range of each data is scaled to [0,1].

[0039] Secondly, due to the harsh sampling environment, environmental influences during the sampling process, or equipment failures, some sample data points may be abnormal. However, the effectiveness of the neural network is extremely dependent on the previous data, and these abnormal points will greatly affect the subsequent prediction results. Therefore, it is necessary to remove the outliers to facilitate subsequent prediction work.

[0040] Finally, due to the difficulty in collecting power system signals, the data points are limited, which is not conducive to deep learning for subsequent power load forecasting. Therefore, the sample data can be decomposed into signals through the local mean decomposition algorithm to expand the sample data and make the prediction more accurate.

[0041] The local mean decomposition algorithm is an adaptive non-stationary signal processing algorithm that decomposes the signal to be analyzed into several product functions. The local mean decomposition algorithm first solves the local mean function and then the envelope estimation function to complete the signal decomposition.

[0042] Compared with other signal decomposition methods, the local mean decomposition algorithm can effectively reduce the number of iterations. In addition, the method does not use cubic splines for interpolation, but uses the sliding average method for envelope estimation. This can be more faithful to the original signal, can completely retain the original information, and make the prediction results more accurate.

[0043] The sample data set after local mean decomposition is divided into training set and test set in a ratio of 8:2 for the subsequent training and testing of the neural network model.

[0044] Step 2: Build a preliminary feature extraction module to extract preliminary features.

[0045] As shown in the network structure of FIG2 , the preliminary feature extraction unit is divided into two paths, each path includes a feature extractor, and each feature extractor is composed of three layers of depth-separable convolution.

[0046] Compared with conventional convolution, depth-wise separable convolution greatly reduces the number of parameters while achieving feature extraction, which shortens the computing cost and time. Depth-wise separable convolution is mainly divided into channel-by-channel convolution DC and point-by-point convolution PC.

[0047] In the depthwise separable convolution, one convolution kernel of DC is responsible for one channel, and the number of channels of the generated feature map is consistent with the number of input channels; the convolution kernel size of PC is 1×1×M, where M is the number of channels of the previous layer. The feature map of the previous layer is weightedly combined in the depth direction to generate a new feature map. The number of parameters used is small, the features obtained are more comprehensive, and effective features can be obtained in a shorter time.

[0048] The two feature extractors have the same structure. The input x0 passes through the two feature extractors to obtain features F1 and F2 respectively. The features F1 and F2 are added at the corresponding positions to obtain the preliminary feature C1, as described in the following formula:

[0049] in, Indicates element-by-element addition at corresponding positions, specifically implemented as: F1 = [f11, f12, f13, ..., f1p], F2 = [f21, f22, f23, ..., f2p],

[0050] Then: C1=[f11+f21,f12+f22,f13+f23,…,f1p+f2p].

[0051] Step 3: Through the residual connection, the preliminary feature C1 is concatenated with the input x0, input into the prediction unit, and the prediction result is obtained.

[0052] The prediction unit consists of 4 layers of BiLSTM. The last layer of BiLSTM is an input-output fully connected network to obtain the prediction results of power load.

[0053] Compared with the long short-term memory network LSTM, BiLSTM can better capture bidirectional semantic dependencies. A BiLSTM is composed of two long short-term memory networks LSTM, which capture the features of the input sequence from the forward and reverse directions respectively. One LSTM processes the input sequence forward; the other LSTM processes the sequence backward. After processing, the outputs of the two LSTMs are concatenated. Each LSTM consists of an input gate it 、Forget Gate t , candidate gate C t and output gate o t It selectively retains important information through gated states and effectively learns time series data. The specific calculation formula is as follows: t =σ(ω f ·[h (t-1) , x t ]+b f ) i t =σ(ω i ·[h (t-1) , x t ]+b i ) o t =σ(ω o ·[h (t-1) , x t ]+b o ) C t =tanh(ω c ·[h (t-1) , x t ]+b c ) h t =o t tanh(C t )

[0054] Among them, x t Input data for the current time t, is the current input state, C t is the current unit state, i.e., the candidate gate, C t-1 is the state of the unit at the last moment, h t is the output of the unit at this moment, h t-1 is the output of the unit at the previous moment. t With h t-1 The two matrices are directly concatenated together as input, and the corresponding weights ω f 、ω i 、ω o 、ω c Multiply, and add the corresponding bias b f 、b i 、b o 、b c The forget gate is used to calculate the last unit state C. t-1 Perform selective forgetting and output the historical information of memory. The input gate is sensitive to the current input state. Selective input is performed, and the output is the current input information. The sum of the memory of historical information and the current input information is the current unit state C tThe output gate selectively outputs the current unit state to obtain the final output h t .

[0055] Only after all time steps are calculated, the forward LSTM obtains a result vector after 6 time steps; the reverse LSTM also obtains another result after 6 time steps. These two result vectors are concatenated to obtain the final BiLSTM output result, which is calculated as follows:

[0056] in, is the forward LSTM output, is the reverse LSTM output, and t1 is the number of layers.

[0057] Step 4: Train the system using the training set and then test the power load prediction effect using the test set.

[0058] Two fully connected layers are connected after BiLSTM to further fuse features and output power load forecast results. The back propagation algorithm is used to perform adaptive parameter adjustment to train the network. The trained network is then verified using a test set. When the loss value falls below the set threshold, training is stopped.

[0059] Example:

[0060] The experimental environment here is as follows: CPU Intel Core i5 13@2.20GHz, GPU NVIDIA RTX3060 12GB, operating system Windows 10.

[0061] The traditional LSTM network, LSTM-Attention combined with the attention mechanism, multi-layer perceptron MLP and support vector machine SVR were used for comparative experiments. The above networks are all commonly used existing technologies for power load. The LSTM network and LSTM-Attention combined with the attention mechanism are currently the most widely used power forecasting methods.

[0062] The experimental results are presented using the mean squared error (MSE) (the average of the absolute squared errors between the predicted and true values), the mean absolute error (MAE) (the average of the absolute errors between the predicted and true values), and the mean absolute percentage error (MAPE) (this metric is sensitive to relative errors and does not change due to global scaling of the target variable). Table 1 shows that the present invention outperforms the other methods in all aspects of error performance. The mean absolute percentage error (MAPE) is the lowest of all methods, at 0.0818. The mean squared error (MSE) of the MLP method is the highest, at 0.5893, while the present invention reduces it to 0.1558, enabling accurate power load forecasting.

[0063] Table 1 Experimental results

[0064] FIG3 is an experimental curve diagram of the present embodiment, showing the curve fluctuations of the prediction results of the present invention and the actual power load at 52 moments. It can be seen from the figure that the prediction results of the present invention are always close to the actual values, further demonstrating the stability and effectiveness of the experimental effect of the present invention, and realizing the ability to accurately predict the power load.

[0065] As shown in FIG4 , another aspect of the present invention provides a power load forecasting device based on a bidirectional long short-term memory network, comprising a data module, a forecasting module and an output module.

[0066] The data module is used to obtain power data and perform preprocessing, normalize the obtained power data, eliminate outlier points and perform signal decomposition to avoid the impact of defects in the data collected by the power system on the prediction performance, making the prediction results more accurate.

[0067] The prediction module includes a preliminary feature extraction submodule and a prediction submodule, which is used to input the data preprocessed by the data module as input, perform power load prediction, and output power load prediction results.

[0068] The preliminary feature extraction submodule extracts preliminary features from the preprocessed power data through two identical feature extractors; the feature extractors are composed of depthwise separable convolutions, and the outputs of the two feature extractors are added together at corresponding positions to form preliminary features.

[0069] The input of the prediction submodule is composed of the input of the preliminary feature extraction submodule and the preliminary features through a residual connection, and then the subsequent feature extraction is performed through BiLSTM and multi-layer perceptron MLP, and the power load prediction result is output.

[0070] The output module is used to transmit the power load forecast results to the display screen via Bluetooth for display. The user can see the forecast results and the corresponding actual detection data through the display screen to complete human-computer interaction.

[0071] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A power load forecasting method based on a bidirectional long short-term memory network, characterized in that: The following steps are involved: Step 1: Obtain power data, preprocess it, and divide it into a training set and a test set; Step 2: construct a preliminary feature extraction unit to extract preliminary features; Step 3: Through residual connection, the preliminary features are concatenated with the input, input into the prediction unit, and the prediction results are obtained; Step 4: Train with the training set and use the test set to test the power load prediction effect.

2. The power load forecasting method based on bidirectional long short-term memory network according to claim 1 is characterized in that: The pre-processing operation in step 1 specifically includes the following processes: First, the acquired power data is normalized and scaled to the same range; Secondly, remove outlier points in the power data; Finally, the power data is decomposed into signals through the signal decomposition algorithm, and the sample data is expanded to make the prediction more accurate.

3. The power load forecasting method based on bidirectional long short-term memory network according to claim 2 is characterized in that: The signal decomposition algorithm adopts a local mean decomposition algorithm.

4. The power load forecasting method based on bidirectional long short-term memory network according to claim 1 is characterized in that: The preliminary feature extraction unit in step 2 is specifically: The preliminary feature extraction unit is divided into two paths, each path includes a feature extractor composed of a plurality of depth-separable convolutions to perform shallow feature extraction, and the structures of the two feature extractors are the same; The features extracted by the two feature extractors are added element by element to obtain preliminary features.

5. The power load forecasting method based on bidirectional long short-term memory network according to claim 1 is characterized in that: The prediction unit in step 3 is specifically: The prediction unit includes an n-layer bidirectional long short-term memory network BiLSTM, and the output of the last layer of BiLSTM is used as the input of the fully connected network, and passes through a multi-layer perceptron MLP composed of m layers of fully connected networks to output the prediction result of the power load.

6. A power load forecasting device based on a bidirectional long short-term memory network, used to implement the power load forecasting method according to any one of claims 1 to 5, characterized in that: Includes data module, prediction module and output module; The data module is used to obtain power data and perform preprocessing; The prediction module is used to take the data input preprocessed by the data module as input, perform power load prediction, and output the power load prediction result; The output module is used to display the power load forecast result on the display screen to complete human-computer interaction.

7. The power load forecasting device based on bidirectional long short-term memory network according to claim 6 is characterized in that: The prediction module includes a preliminary feature extraction submodule and a prediction submodule; The preliminary feature extraction submodule extracts preliminary features from the preprocessed power data through two identical feature extractors; The input of the prediction submodule is composed of the input of the preliminary feature extraction submodule and the preliminary features through a residual connection, and then the subsequent feature extraction is performed through BiLSTM and a multi-layer perceptron, and the power load prediction result is output.

Images