Decomposition prediction fractional-order PID reinforcement learning for short-term smart generation control of integrated energy systems

With the continuous development of integrated energy systems (IESs), various distributed power is continuously connected to IESs. Uncertainty and volatility of renewable energy could increase power deviations of power systems and put pressure on the frequency control of the grid. A decomposition prediction fractional-order PID reinforcement learning (DPFOPIDRL) algorithm is proposed to reduce frequency deviations of IES and improve power quality in this study. The DPFOPIDRL-based unified timescale intelligent generation controller, which sends regulation commands to the automatic generation control every four seconds, continuously collects time series signals of frequency deviations; and then, the time series signals are predicted after modal decomposition. The DPFOPIDRL applies state-action-reward-state-action to control prediction signals with intense fluctuations, and fractional order proportional-integral-derivative to control prediction signals with gentle fluctuations. The DPFOPIDRL, proportional-integral-derivative, Q learning, and sliding mode are compared in four cases in an equivalent simplified and complex IES based on IEEE 39-node-Belgium 20-node system. Results under complex IES show that the frequency deviation and total generation cost of the DPFOPIDRL are at least 35.66% and 16.13% smaller than PID, Q learning, and sliding mode, respectively.