ANN-BASED SUPERVISORY CONTROL OF VEHICLE-TO-GRID TECHNOLOGY IN A MICROGRID USING DC FAST CHARGING ARCHITECTURE

Abstract

Electric vehicle (EV) batteries can act as energy storage in micro-grids by charging when there is surplus generation (grid-to-vehicle, G2V) and returning energy when it is needed (vehicle-to-grid, V2G). Mohammed Shakeel and Malik proposed a DC fast-charging station for V2G in a grid-connected microgrid, in which four EVs and a 50 kW photovoltaic array share a 1.5 kV DC bus that is connected to the micro-grid through a 250 kVA grid-connected inverter with an LCL filter. In that work the battery chargers and the inverter use fixed-gain proportional–integral (PI) controllers, which produce DC-bus voltage spikes of 40–65 V and battery-current overshoot of about 11 % whenever an EV changes mode. This thesis replaces every PI regulator of the station — the constant-current loops of the off-board chargers and the dq current and DC-voltage loops of the inverter — with an artificial neural network (ANN) controller. Each ANN is a 3–10–1 multilayer perceptron whose inputs are the error, its integral and its derivative, and it slots into the original control structure without changing the feed-forward, decoupling or modulation stages. A two-stage training method is proposed: the networks are first fitted to the PI law with the Levenberg–Marquardt algorithm, giving a stable starting point, and are then optimised in closed loop with an evolution strategy on a station model that includes computation delay and sensor noise. A PI controller re-tuned on the same objective is used as an additional benchmark. Simulation studies covering the base-paper V2G–G2V scenario, PV variation, four simultaneous EVs, parameter variation, grid voltage sag, power quality and sensor noise show that, compared with the basepaper PI, the ANN reduces the DC-bus voltage deviation by 20–29 %, roughly halves the DC-bus settling time, lowers the ITAE by 64 %, cuts battery-current overshoot to 2–5 % and reduces inverter modulation noise by 60 %. The optimised PI benchmark achieves nearly the same dynamics, so the benefit specific to the ANN structure is modest. The ANN also gives higher grid-current THD than the base PI (1.08 % against 0.74 % under base-paper conditions, but 5.82 % against 4.24 % with dead time) and loses part of its advantage under combined parameter variation. These limitations, and the remedies proposed for them, are discussed.