A quantitative analysis of machine learning models for short- term generation forecasting

Abstract

Accurate short-term electrical generation forecasting is of great importance to grid stability and reliability as it improves the performance of power scheduling, dispatching scheduling, financial planning, and energy management schemes. Short-term electrical demand is non-stationary and nonlinear, rendering traditional statistical forecasting approaches which were built on linear correlation and stationarity ineffective. It is vital, that an adequate power planning system with a high level of forecasting accuracy be implemented as we face the power demand of the twenty-first century. Furthermore, with the rapid advancement in artificial intelligence and machine learning technologies, bundled with the increase in the availability of data and the enhancement of data computing devices, science has witnessed remarkable achievements in fields such as natural language processing, medicine, communication, and transport. In this research, a comparative and quantitative analysis was carried out using 3 machine learning algorithms for short-term generation forecasting as a means of improving prediction accuracy, in the case study of Song Loulou Hydro-power plant. The Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) values were used to quantify the accuracy of the algorithms. Furthermore, the ARIMA traditional statistical model was used as a validation and comparative approach to test if the added complexity of using machine learning algorithms was an indication of better performances. Python programming language and its libraries on the Jupyter notebook IDE were used to analyze the data. Results from this research will improve grid stability and reliability of the southern interconnected grid of Cameroon which is supplied by the Song Loulou hydro-power plant. The experimental results clearly showed the advantage these modern techniques have over the traditional statistical approach. The machine learning Recurrent Neural Network (RNN) using Long Short-Term Memory (LSTM) produced the best result with an RMSE value of 10.8 MW and a MAPE value of 3.4% as compared to the ARIMA model which RMSE value was 32.9 MW and MAPE value was 10.6%. In addition, the other machine learning algorithms also outperformed the ARIMA algorithm with the Random Forest model and the Regression model having an RMSE value of 15.3 MW and 15.8 MW respectively. In conclusion, integrating machine learning techniques as a tool in the power sector will greatly improve the performance of the grid.