Very short term load forecasting of a distribution system with high PV penetration

Abstract

A battery energy storage system (BESS) is an available solution for utilities to deal with intermittency issues resulting from renewable energy resources. A BESS needs to have a control algorithm to provide a very good estimation of the load on the grid at each time step. A short-term load forecast (STLF) is necessary for efficient and optimized control of BESSs that are connected to the grid. In this work, two parallel-series techniques for load forecasting are proposed to optimize the performance of a grid-scale BESS (1 MW, 1.1 kWh) in 15-min steps within a moving 24-h window. In both techniques, a complex-valued neural network (CVNN) is used for parallel forecasting. The parallel component is based on the search for similar days of historical data that have a weekly index comparable to the forecast day. For series forecasting, historical data of each day is used within a moving forecast window by CVNN along with the spline method. For both techniques, parallel forecasting is mixed with series forecasting by an adjustment coefficient. Both techniques are tested on a set of real data for a grid with high PV penetration, and the obtained results are compared.