An Innovative Adaptive Threshold-Based BESS Controller Utilizing Deep Learning Forecast for Peak Demand Reductions

Battery-based energy storage systems (BESS) can be installed and controlled by a control strategy to reduce daily peak demands. Majority of the existing controllers for BESS is developed and studied on a simulation platform. Hence, an innovative adaptive threshold-based controller is developed and implemented on a real 200 kWh BESS installed on a university campus in Malaysia to reduce daily peak demands. This controller is developed on the free, open-source Node-RED platform using a deep learning-based one-dimensional convolution neural network (1D-CNN) to forecast the load profile for one day ahead. The controller uses the forecasted load profile to define a threshold or specific power demand level above which the BESS will start discharging power to meet the demand. The threshold is actively adjusted according to the actual and forecasted power demand of the customer as well as preceding peak of the grid power. The performance of the controller is initially assessed through six months of comprehensive simulations, where its daily peak reduction factor (K<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">pdr</sub>) and monthly peak reduction failure rate (η<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">failure</sub>) are compared with that of two fixed-threshold controllers, namely forecasted threshold and historical threshold-based controllers, and another two advanced controllers, namely active and fuzzy controllers. The adaptive threshold-based controller achieves an average K<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">pdr</sub> of 41.62% and η<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">failure</sub> of just 16.55%, which are better than that of other four controllers. The controller is implemented to BESS in a university building to evaluate its practical performance over 21 days under a real operating condition. The controller achieves an average K<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">pdr</sub> of 49.45% with a η<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">failure</sub> of just 4.76%, which are better than the simulation results.