동적효율 향상을 위한 Neural Network 기반의 새로운 태양광

Abstract

This paper deals with a novel Neural Network (NN)-based P&O MPPT control technique for maximizing dynamic efficiency of solar power generation. To maximize the efficiency of photovoltaic (PV) panels, which have low energy densities and nonlinear characteristics where output is determined by various environmental variables such as irradiance and temperature, high-performance Maximum Power Point Tracking (MPPT) control is crucial. Conventional Hill Climbing (HC)-based MPPT control techniques, such as Perturb & Observation (P&O) MPPT and Incremental Conductance (InC) MPPT, are the most representative. Their simple algorithms facilitate easy application and relatively stable performance, making them widely used in PV MPPT control. However, these HC-based MPPT control techniques struggle to handle the nonlinear characteristics of PV output, which makes interpretation of input-output correlations challenging under rapidly fluctuating irradiance. Unlike conventional HC-based MPPT control, AI-based MPPT control can appropriately handle the nonlinear characteristics of PV output, which makes it difficult to interpret input-output correlations in situations where irradiance fluctuates rapidly. This is why it has attracted significant attention in PV MPPT control. Research on various AI-based MPPT control techniques has been conducted to date, and multiple AI techniques such as Fuzzy Logic (FL), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO), Support Vector Machine(SVM), Neural Network (NN) have been applied to numerous PV MPPT studies, and NN is a fundamental technology underpinning modern AI that have been extensively utilized in MPPT applications. However, in most NN-based MPPT control techniques, the NN output becomes the control variable (Duty or VREF). If the NN output is directly reflected in the control variable, the learning error directly leads to MPPT control error, significantly increasing the impact of the NN learning state on MPPT performance. Therefore, to reduce the impact of NN learning error and enhance control stability, an approach that separates the NN output from the control variable is necessary. In this paper, a novel NN-based P&O MPPT control technique that separates the NN output from the control variables is proposed. Instead of directly inputting the NN output as a control variable, we propose a PV-OP Estimator that outputs a score that identifies the current PV output operating point. This PV-OP Estimator utilizes the basic structure of a Feedforward Neural Network (FNN) with eight physical quantities (duty, PV panel output voltage, current, and power, converter output voltage, current, and power, and irradiance) as input features and two hidden layers, resulting in a score as the final output. Applying the proposed PV-OP Estimator to an MPPT control algorithm and performing irradiance step response and EN50530 dynamic tests, we achieved improved performance compared to existing methods. The proposed FNN-based MPPT control technique requires five sensors, necessitating minimizing the number of sensors for commercialization. Therefore, a Time-Delay Neural Network (TDNN) MPPT control technique that achieves comparable or better performance with only two sensors (solar panel output voltage and current) is also proposed. Compared to the conventional FNN-based control technique, the TDNN-based MPPT control technique demonstrated comparable or better MPP tracking performance despite reducing the number of sensors by less than half. Computer simulations and experiments were conducted to demonstrate the validity of the proposed FNN-based P&O MPPT control technique and the TDNN-based P&O MPPT control technique.