Intelligent Systems with Applications (Sep 2023)
An adjustable machine learning gradient boosting-based controller for PV applications
Abstract
Traditional Proportional, Integral, and Derivative (PID) controllers are employed in most industrial control applications because they are simple and easy to implement. However, they are fallible and exhibit poor performance in complicated, non-linear, time-delayed systems, and noisy feedback loops. In photovoltaic systems applications such as maximum power point (MPP) tracking, PI controller performance in low irradiance levels, varying load conditions, and partial shading scenarios is considerably inefficient in tracking the corresponding reference current, which generates MPP and causes considerable steady-state error. Machine learning can be used with selective input/output data from the PI controller to build a model that has the ability to adjust in case of steady-state error. Thus, an adjustable Machine Learning Gradient Boosting-based (MLGB) controller with a PV module and a DC-DC boost converter is proposed in this work. To create the raw dataset, a PI controller was simulated, and the input/output signals were recorded. Data preprocessing, utilizing feature engineering and Shapley Additive Explanations (SHAP) values, was used to explain the dynamic behavior of each feature, the inter-feature dependability, and their significance with reference to the model output. Hyperparameters were tuned with cross-validation while the model was being created using the CatBoost method. To assess the adjustment of each feature to minimize the error, particularly for lower irradiance values, varying load conditions, and partial shading, a regression model was built. Using the MATLAB Simulink Environment, a complete system controlled by both the PI and the MLGB controller was developed and tested. A random irradiance level fluctuation profile, along with varying load and partial shading scenarios, was used to compare three controllers, namely the PI, conventional MLGB, and adjusted MLGB controllers, to see how they respond to rapidly changing environmental conditions. It has been proven through a thorough simulation analysis that the novel adjustable MLGB controller exhibits good tracking performance to follow the DC-DC boost converter's inductor current. When compared to the traditional PI controller, the new controller exhibits better results in terms of steady-state behavior and transient responsiveness in general, with a much lower mean (3.059E-03), median (1.908E-03), and RMS value (2.168E-02) of the signal error under various conditions.