Short-term load forecasting of industrial park considering time correlation under data-driven

Abstract

The load characteristics of the park have strong fluctuation and nonlinearity due to the comprehensive power supply capacity, weather and holidays. This paper proposes a data-driven short-term load forecasting model, taking into account meteorological and temporal features of industrial parks, that takes into account time dependence. In the process of prediction, we build a convolutional neural network based on PyTorch framework, which combines short term memory network and autoregressive model. We construct a convolutional neural network based on PyTorch framework, a combination of short-term memory network and autoregressive model, in order to forecast. By unifying the benefits of CNN and LSTM, CNN-LSTM can extract features, seize long-term dependencies, and minimize overfitting. Therefore, we first extract both spatial and time series features by combining CNN-LSTM networks. Feature extraction modules, such as the CNN convolution layer and pooling layer, are employed to acquire data features. Temporal correlation properties of data can be mined through Long Short-Term Memory (LSTM) networks. Nonlinear time series prediction is achieved by employing Autoregressive Integrated Moving Average (ARIMA) to extract periodic and trend factors from the time series data. Finally, the prediction results are obtained according to the inverse error combined weights. Based on the load data of an industrial park in Liaoning Province, this paper carries out simulation experiments. The CNN-LSTM-ARIMA model has been demonstrated to be a successful means of enhancing the accuracy of load forecasting in industrial parks, when compared to LSTM and CNN-LSTM models. The practical engineering application of the prediction model holds a certain worth.