Electricity price forecast based on the STL-TCN-NBEATS model

• DOI: 10.1016/j.heliyon.2023.e13029
• Publication Date: 2023-01-14T16:07:39Z
• Authors: Biao Zhang; Chao Song; Xuchu Jiang; Ying Li
• Keywords: Interpretability; Autoregressive–moving-average model; Robustness

Taking long-term high-frequency electricity price data as the research content, this paper proposes seasonal and trend decomposition using loess-temporal convolutional network-neural basis expansion analysis for an interpretable time series forecasting (STL-TCN-NBEATS) model to solve problems of low forecast accuracy caused by high volatility, frequency nonlinearity poor interpretability deep learning model. By comparing effects temporal network-long short-term memory (TCN-LSTM), LSTM other models, main conclusions are follows: (1) The hybrid model, STL-TCN-NBEATS, selected in can effectively problem after reasonable selection parameters. evaluation indexes root mean square error (RMSE) absolute percentage (MAPE) were 3.7441 4.5044, respectively, which 3.1416 2.1336 lower than those second-best (TCN-LSTM). Compared with autoregressive integrated moving average (ARIMA), was improved approximately 49.18% 60.35% (MAPE). (2) STL-TCN-NBEATS has better feature extraction ability, so it obtain higher accuracy. Since construction TCN introduces extended causal convolution residual blocks, network processing ability robustness large sample series. Moreover, NBEATS structure enables be trained quickly, experimental results verify effectiveness method. (3) not only precision but also some interpretability. decomposing into a term, period term remainder used process that prices according traditional mode.