A hybrid prediction interval model for short-term electric load forecast using Holt-Winters and Gate Recurrent Unit

Abstract

Short-term load forecasting (STLF) is essential for power system planning and operation decision-making. Researchers in this field have devoted themselves to building point prediction models in recent years. However, traditional point forecasting treats the forecast outcome as a deterministic variable. Due to the inherent complexity, volatility, and instability of power load, deviations in traditional point forecasting methods are inevitable and significant. Considering the difficulty of accurate point prediction, interval prediction can tolerate increased uncertainty and provide more information for actual operation decision-making. In order to solve this problem, we proposed a novel hybrid model based on Holt-Winters (HW) method and gated recurrent unit (GRU) network for short-term load interval forecasting. First, considering the factors affecting the load trend, the weather type, date type, and historical load data are processed and normalized. Then, the raw data is decomposed into linear components and nonlinear residuals using a moving average (MA) filter. Next, the HW method establishes a linear prediction model to predict the linear component. Finally, an interval prediction model based on gated recurrent unit neural network is established by taking linear prediction results, nonlinear residuals, weather types, date types, and original data as inputs. In order to verify the performance of the proposed hybrid method, based on the data set provided by the 2022 "Teddy Cup" data mining challenge, taking 15-minute load data as an example, the traditional model and the advanced interval prediction model were compared. The results show that the proposed model is a high-quality method. Compared with models built by other methods, it has a higher prediction interval coverage probability and a narrower prediction interval width.