A novel ensemble model for long-term forecasting of wind and hydro power generation

Abstract

Power generation scenario modelling has become an integral part of long-term planning in power system due to high penetration of variable renewable energy. It requires accurate estimates of power generation from different resources to find cost-optimal mix of generations. Predicting the generation of weather dependent renewables in long-term is not feasible but an adaptive long-term forecasting model based on univariate time-series analysis can provide the solution. Therefore, an effort has been made through this paper to provide accurate medium to long-term forecasts (a week-ahead to a year-ahead) for wind and hydro power generation using a novel ensemble forecasting model. The proposed model is devised in three phases; Phase-I develops a hybrid model using ARIMA (Auto Regressive Integrated Moving Average) and Bi-LSTM (Bidirectional Long Short Term Memory) predictions. Phase-II integrates the forecasts of seasonal and off-season generation periods obtained via a Diligent Search Algorithm (DSA). DSA is an innovative algorithm, designed to identify the hidden seasonalities that are responsible for the intermittent behaviour of wind and hydro power generation time-series. Finally, Phase-III facilitates amalgamation of prediction results of Phase-I and Phase-II to build the proposed forecasting model. Results show that MAE (Mean Absolute Error) for wind and hydro power are 1.97% to 5.52% and 2.3% to 6.42% while RMSE (Root Mean Square Error) varies from 2.79% to 7.8% and 2.63% to 8.4% respectively in a week-ahead to a year-ahead scenarios. Since this model is specifically designed for a year-ahead forecasting scenario, its performance can become unstable beyond this time horizon.