Quick MPPT microconverter using a limited general regression neural network with adaptive forgetting

The recent development of photovoltaic electricity generation technology has led to a large number of solar-farms worldwide. Photovoltaic systems are a current source that cannot be used as normal batteries. Therefore, we have to set the voltage properly to obtain maximum power from the photovoltaic. Although there are various techniques for realizing maximum power point tracking, only a few techniques are able to realize a quick response to sudden irradiation changes using conventional circuit design. To construct a quick controller with a conventional circuit configuration, we proposed a model-based maximum power tracking microconverter that learns the relationships among the strength of irradiation, temperature, and the maximum power point tracked by the perturbation and observation method. After the learning, if solar irradiation suddenly changes, and the temperature and strength of solar irradiation is close to one of the learned pattern, it directly sets the recalled reference voltage for the chopper circuit. Using this strategy, the microconverter realizes quick control to large and sudden changes in solar irradiation. Moreover, we need not provide any prior knowledge to the learning machine. However, the learning method employed in the previous attempt to keep learned control skill under a fixed budget. Therefore, there are cases when learning some new instances is omitted to preserve memory. This leads to a problem wherein the converter cannot adapt to changes in the properties of a photovoltaic. To overcome the problem, we introduce a modified learning method for the microconverter, which attempts to forget certain memories to make space to learn new instances. The experimental results show that the new method adapts to the environmental changes well while maintaining its quick-response properties.