Design and Dynamic Performance Evaluation of an LSTM Based Deep Learning Controller for PV Systems Operating in Different Voltage Regions

In this paper, a deep learning Long Short Term Memory (LSTM) controller is designed using multiple optimal PI controllers tuned in different operating regions. It has been previously shown in the literature the impact of climatic conditions and abnormal operating conditions on the power conversion efficiency of photovoltaic (PV) systems. The nonlinear characteristic curve of PV arrays exhibits an additional transient effect that influences the tracking of the Maximum Power Point MPP. The PV power conversion system is characterized by a variable open-loop transient response in the constant current, voltage and power regions, which are subdivisions of the PV array characteristic curve. For the optimal tracking of the reference generated from the MPPT algorithm in these operating regions, an input/output data collection is carried out from the closed-loop system responses using multiple PI controllers tuned in different operating regions. Then, the LSTM controller is tuned using the collected training data constructed from the concatenation of input/output data issued from all operating regions. The dynamic performance evaluation of the deep learning-based LSTM controller for different simulation scenarios, including reference step changes, stair-shaped reference changes and partial shading tracking, shows the high precision and reduced oscillations of the responses issued after using the proposed controller.