A Load Forecasting Method of Power Grid Host Based on SARIMA-GRU Model

With the continuous development of intelligent power grid, how to boost the prediction ability of the future operating mode of information equipment and set the dynamic provision for prediction intervals to adapt to the changes of data are huge challenges for the grid IT operation and maintenance. To solve these problems, this paper proposes a combined time series forecasting model (SARIMA-GRU) based on the traditional Seasonal ARIMA model (SARIMA) and the GRU model in Deep learning, which is the error-fitting model by using the error auto-regressive method to compensate for the prediction result. In order to establish the threshold interval in line with the actual production demand, SARIMA-GRU applies the statistical method and K-nearest neighbor algorithm for global preprocessing, and then divides the non-stationary series into three main components of model: trends, seasonality, and residual terms. By using the corresponding model components to predict, we achieve higher prediction accuracy under the normal operation state. On a real-world power grid dataset, we demonstrate more significant performance improvements over the traditional model ARIMA, SARIMA and combination model, like ARIMA-SVM, and showcase three actual threshold intervals.