Assessing the impact of load forecasting accuracy on battery dispatching strategies with respect to Peak Shaving and Time-of-Use (TOU) applications for industrial consumers

Abstract

Energy Storage Systems will play crucial role in controlling the grid of the future when increased penetration of renewable energy sources will take place. Especially batteries are expected to occupy a considerable share of the total energy storage market by simultaneously providing services to different stakeholders such as energy producers, transmission/distribution operators, residential, commercial and industrial consumers. Nowadays, Peak shaving and Time-of-Use applications are the most common services that standalone battery storage systems can provide to industrial consumers (without integrated PV-systems and/or wind turbines). A big part of the existing literature addressing such applications aims at developing an offline algorithm for optimal battery deployment based on a known load profile (or accurately predicted) without taking into consideration real time conditions. This paper investigates the impact of industrial load forecasting errors on dispatching strategies of battery storage systems on economically driven peak shaving and Time-of-Use applications. An artificial neural network has been developed and used as a prediction model of an industrial load profile. The neural network was trained, validated and tested on historical load data with time resolution of 15 minutes, provided by the local distribution operator of the Belgian electric grid. The performance of the neural network in terms of output-target regression and mean absolute error is 0.833 and 10.02% respectively. Afterwards, a simulation was carried out comparing four different scenarios of peak shaving. The results show that the prediction accuracy of the presented neural network is not competitive enough. Peak shaving based on predicted profiles becomes reliable for lower forecasting errors. For this purpose, further access into the process and types of loads of the user is required in order to come up with a more sophisticated prediction model.