Facing the problem of increasing load fluctuation and green power consumption in power distribution network (PDN) caused by high proportion of distributed photovoltaic access, this article presents a data-driven adaptive load regulation and intelligent optimization method. Firstly, the real operation data of low-voltage stations in Beijing area are cleaned and feature extracted, and the Seq2Seq-Attention photovoltaic power generation prediction model is built to achieve accurate power prediction in the next 6 hours (the test set MAE is 0.18 kW). Then, a reinforcement learning control algorithm based on Deep Deterministic Policy Gradient (DDPG) is designed to dynamically adjust the charging and discharging of energy storage and the adjustable load input in order to stabilize the fluctuation of net load and improve the spontaneous self-use rate. The experiment was carried out on the OpenDSS simulation platform, including many typical scenes such as sunny summer and rainy winter. The results show that, compared with the traditional rule control, the average reduction rate of peak-valley difference is 18.7%, the local absorption rate of photovoltaic is increased to over 89.3%, and the voltage over-limit time is greatly reduced.
Rao et al. (Sun,) studied this question.