Integrating data decomposition and machine learning methods: An empirical proposition and analysis for renewable energy generation forecasting

Abstract

Renewable energy generation (REG) has been vigorously developed under the dual pressure of environmental contamination and fossil fuel energy shortage. Whereas the rapidly increasing REG instigates stochastic volatility may threaten the REG deployment and grid stability. Motivated by the grid flexibility enhancement depending on the expansion of grid infrastructure that belongs to the mid-term management domain, the precise predictions for mid-term, especially monthly REG sourced from diverse renewable energy, are indispensable for power plant configuration, electricity end-uses promotion, and grid expansion to propel the energy system transformation. To this end, aiming at diverse monthly REG datasets characterized by periodicity, nonlinearity, and volatility, the seasonal-trend decomposition procedure based on loess (STL) is initially employed to extract the trend and periodic features of monthly REG datasets, and generate trend, seasonal, and remainder subseries. Based on the decomposed data, long-short term memory (LSTM) is utilized for subseries prediction, and then the projections are integrated to compose the ultimate forecasted results for original REG observations. To validate the efficacy and adaptability of the proposed data-driven STL-LSTM framework, several machine learning methods and autoregressive models are involved in predicting practical monthly electricity generation datasets sourced by solar, wind, hydropower, and geothermal energy in different countries. The forecasted results indicate that the proposed framework is demonstrated with superlative prediction performance and strong adaptability for generation prediction derived from diverse renewable energy. Therefore, the novel STL-LSTM framework may constitute a promising alternative for REG forecasting.