A reinforcement learning based approach for on-line adaptive parameter extraction of photovoltaic array models

摘要

At present, most methods for the fault detection and diagnosis (FDD) of the photovoltaic (PV) array strongly rely on comparing the on-line measured electrical parameters with the modeled reference ones, which are challenging the on-line accuracy and time cost of the parameter extraction for modeling the current-voltage (I-V) curves of the PV array. In this paper, a reinforcement learning (RL) based approach for on-line adaptive parameter extraction of PV array models is proposed. The model parameters, including the ideality factor, series and shunt resistance, and the compensated irradiance for the uncalibrated pyranometer, are extracted. Corresponding environmental states, actions, rewards, and the entire framework for the on-line adaptive parameter extraction are reasonably designed and investigated. The annual experimental results verify that the proposed RL-based approach can obtain higher on-line accuracy for modeling the I-V curve of PV array with fast extraction speed, compared with the conventional meta-heuristic-based approach and the analytical approach for parameter extraction. The annual experimental results reveal that the proposed approach can guarantee the 50% probability for obtaining the root mean square error (RMSE) less than 0.1, and 90% probability for obtaining the RMSE less than 0.25. The average computational time cost of the proposed approach is approximate 38.12 ms. In addition, the annual trend of extracted model parameters is analyzed. The annual results also show that the series and shunt resistance have the inverse seasonal trend. Besides, the measurement error of the pyranometer can be identified statistically. The proposed RL-based approach can also be integrated with the presented on-line FDD method, which realizes the on-line training of RL agents and the FDD of PV array simultaneously.