Short-term power load forecasting method based on improved AE-LSTM

Abstract

The application discloses a kind of short-term electric power load prediction methods based on improved AE-LSTM, it is related to electric power system data analysis and deep learning technical field.The method includes: obtaining historical electric power load time series data and carrying out Z-score standardization preprocessing;Using denoising autoencoder to the data after preprocessing is carried out feature extraction and noise reduction processing;The extracted feature is input into residual BiLSTM-temporal-spatial attention hybrid model, and the future short-term electric power load prediction result is output.The application improves input feature quality by denoising autoencoder, solves deep network gradient vanishing problem using residual connection, and captures the temporal-spatial correlation of load through the temporal-spatial attention module.Compared with the basic LSTM model, the MAPE of the application reduces by 76.1%, R2 increases to 0.9924, the prediction accuracy is greatly improved, and is suitable for smart grid short-term electric power load prediction scene.

Description

Technical Field

[0001] This invention relates to the field of power system data analysis and deep learning technology, and in particular to a short-term power load forecasting method based on an improved AE-LSTM. Background Technology

[0002] Electricity load forecasting is a core component of smart grid dispatching, planning, and operation. Accurate short-term load forecasting can effectively improve grid operating efficiency, reduce generation costs, and ensure the safe and stable operation of the power system. Traditional LSTM models, when processing time-series electricity load data, can only capture local features in the time dimension, making it difficult to fully explore the spatiotemporal correlations of electricity load. Furthermore, deep networks are prone to the vanishing gradient problem, leading to insufficient prediction accuracy. Simultaneously, electricity load data inevitably contains random noise, which, if directly used for model training, will further degrade the model's predictive performance. Therefore, targeted improvements to traditional LSTM models are needed to enhance the accuracy of short-term electricity load forecasting. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a short-term power load forecasting method based on an improved AE-LSTM. By combining a denoising autoencoder, residual BiLSTM and a spatiotemporal attention mechanism, the method effectively extracts deep features of power load, captures spatiotemporal correlations, and significantly improves the forecasting accuracy of short-term power load.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A short-term power load forecasting method based on an improved AE-LSTM includes the following steps:

[0006] S1: Acquire historical power load time-series data and perform standardized preprocessing on the data;

[0007] S2: A denoising autoencoder is used to perform feature extraction and noise reduction on the preprocessed load data;

[0008] S3: Input the extracted features into the improved LSTM prediction model and output the future short-term power load prediction results;

[0009] The improved LSTM prediction model is a residual BiLSTM-spatiotemporal attention hybrid model.

[0010] Furthermore, the denoising autoencoder includes an input layer, a hidden layer, and an output layer. It extracts deep features from power load data through unsupervised learning while removing random noise from the original data, thereby improving the quality of the input features.

[0011] Furthermore, the residual BiLSTM-spatiotemporal attention hybrid model introduces a residual connection structure into the bidirectional LSTM network to solve the gradient vanishing problem in deep networks and enhance the model's feature representation ability; and adds a spatiotemporal attention module before the network output layer to adaptively assign different weights to different time steps and spatial features, thereby improving the model's ability to capture key features.

[0012] Furthermore, the standardization preprocessing in step S1 adopts the Z-score standardization method to normalize the original load data to an interval with a mean of 0 and a variance of 1, thereby eliminating the influence of dimensions and accelerating the model convergence speed.

[0013] The beneficial effects of this invention are as follows: Compared with the basic LSTM model, the improved prediction model of this invention reduces MSE by 92.8%, MAE by 75.4%, and MAPE by 76.1%, and increases R² to 0.9924. The prediction accuracy and fitting effect are greatly improved, which can meet the practical application needs of short-term power load forecasting in smart grids. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to specific embodiments.

[0015] In this embodiment, the experiment is conducted using the UCI public electricity load dataset, and the specific steps are as follows:

[0016] 1. Data preprocessing: Obtain historical hourly power load data and normalize the data to an interval with a mean of 0 and a variance of 1 using the Z-score standardization method;

[0017] 2. Denoising Autoencoder Training: Construct a denoising autoencoder containing an input layer, a hidden layer, and an output layer. Set the number of neurons in the hidden layer to 64, use the Adam optimizer, a learning rate of 0.001, and 50 training epochs. Perform feature extraction and denoising on the preprocessed payload data.

[0018] 3. Prediction model training: Construct a residual BiLSTM-spatiotemporal attention hybrid model, set the sequence length to 24 (using data from the past 24 hours to predict the load for the next hour), and the training parameters are: training epochs=100, batch size=32, the optimizer is Adam, and the learning rate is 0.001.

[0019] 4. Result Validation: The processed dataset is divided into training and testing sets in an 8:2 ratio to train and test the model.

[0020] Experimental results show that, compared with the basic LSTM model, the improved model of this invention reduces the MSE from 0.055929 to 0.004060, the MAE from 0.183431 to 0.045247, the MAPE from 1.64% to 0.39%, and the R² from 0.8953 to 0.9924 on the test set, significantly improving prediction accuracy and fitting performance.

Claims

1. A short-term power load forecasting method based on an improved AE-LSTM, characterized in that, Includes the following steps: S1: Acquire historical power load time-series data and perform standardized preprocessing on the data; S2: A denoising autoencoder is used to perform feature extraction and noise reduction on the preprocessed load data; S3: Input the extracted features into the improved LSTM prediction model and output the future short-term power load prediction results; The improved LSTM prediction model is a residual BiLSTM-spatiotemporal attention hybrid model.

2. The short-term power load forecasting method according to claim 1, characterized in that, The denoising autoencoder includes an input layer, a hidden layer, and an output layer. It extracts deep features from power load data through unsupervised learning while removing random noise from the original data.

3. The short-term power load forecasting method according to claim 1, characterized in that, The residual BiLSTM-spatiotemporal attention hybrid model introduces a residual connection structure into a bidirectional LSTM network and adds a spatiotemporal attention module before the network output layer to adaptively capture the temporal and spatial correlations of power load.

4. The short-term power load forecasting method according to claim 1, characterized in that, The standardization preprocessing in step S1 uses the Z-score standardization method to normalize the original load data to an interval with a mean of 0 and a variance of 1.