Short-term load forecasting using detrend singular spectrum fluctuation analysis

Abstract

The accuracy of short-term load forecasting (STLF) is susceptible to the complex components of original time series. Conventional data decomposition algorithms, such as singular spectrum analysis (SSA), cannot determine the complex components and thus fails to improve the accuracy of STLF significantly. Given this, this paper proposes a novel decomposition algorithm, namely detrend singular spectrum fluctuation analysis (DSSFA), to improve the accuracy of STLF. The novel algorithm extracts the trend and periodic components from original series using linear and sine function, respectively. The rest series are decomposed by SSA and the long-rang correlation components without white noise are obtained using fluctuation analysis. Long short-term memory (LSTM) is the model used for forecasting the long-rang correlation components. Combining the forecasts of trend, periodic, and long-rang correlation components, we receive the final forecasting results of DSSFA-LSTM. In our experiments, we design a case study with the most recent load of four exit points within natural gas pipeline. The results show that DSSFA outperforms SSA in improving the performance of forecasting models, when dealing with the short-term load series with high complexity. In Oinofyta, DSSFA-LSTM perfectly fit the real load series and its R2 is 1.8 times higher than that of LSTM.