Optimizing Energy Storage System (ESS) for Charging and Discharging Operations via Machine Learning

Abstract

Powerpal like other energy companies faces the problem of effective energy distribution to its clients. Currently Energy Storage System (ESS) deployed by Powerpal is being charged at midnight to its full capacity (325kWh) for next day operations. Following this method of charging does not guarantee optimal results in terms of efficiency and profit for Powerpal. The solution to this problem is the topic of research for this thesis paper.

This thesis paper examines whether machine learning can enhance the efficiency and cost effectiveness of charging and discharging operations for ESS at Powerpal, comparing it to the current ESS model utilized by Powerpal. Powerpal has been keeping track of their operations by recording time series data. This provided an opportunity to deploy machine learning particularly Long short-term memory (LSTM) on their historical data. This machine learning model forecasted energy demand (from electric vehicles and Smedvigkvartalet) and energy generation (from solar and the electric power grid) facilitating informed decisions on ESS charging and discharging.

Two separate LSTM methods i.e. Univariate and Multivariate were trained on dataset from Powerpal. In order for the comparison between these two models to be fair, hyperparameters i.e., learning rate, sequence length, epochs, optimizer, hidden state size and activation function in dense layer were kept constant for the two models. The percentage change between the values predicted my Univariate in comparison to actual data was i.e., maximum difference of -9.2 % and minimum difference of -1.2 . Whereas in the case of Multivariate value prediction in comparison to actual data, the percentage difference was i.e., maximum difference of -47.2 % and minimum difference of -13.9 %. Univariate was better at making accurate prediction in comparison to Multivariate. After predicting the supply and demand of energy from both these models for April 2024, cost analysis was done i.e., which model resulted in more money saved by Powerpal for ESS operations. Because of more accurate predictions made by Univariate LSTM, Powerpal saved more money in comparison to Multivariate LSTM for its ESS operations for the month of April 2024. With more accuracy in expected supply of energy from solar and grid predictions, the utilization of ESS for charging and discharging was less, hence decreasing the operational cost of running ESS, that is why Univariate resulted in more money saved for ESS operations.

The results showed the potential of LSTM machine learning model in making predictions on Powerpal dataset. Powerpal can utilise these results to include machine learning in their daily operations for ESS.