A Very-Short-Term Online PV Power Prediction Model Based on RAN With Secondary Dynamic Adjustment

Abstract

Photovoltaic (PV) power is progressively being subsumed into power grids. As a consequence, reliable PV power forecasting has become essential in order to ensure the optimal functioning of the power grid. Neural networks remain the dominant prediction model utilized. Conventional neural network forecasting models are wholly dependent upon offline data. Subsequent to offline training, no further structural adjustments can be made during the forecasting process, which therefore they fail to cater for PV power supply fluctuations which are fundamentally dynamic. To address this failing, this paper proposes a very-short-term online prediction model based on a resource-allocating network (RAN) incorporating a secondary dynamic adjustment. The RAN is initially trained offline to obtain a basic forecasting model. Thereafter, in an online prediction process, those samples with large prediction errors which exceed a preset value are recorded in a specific buffer. When the set conditions are triggered, the secondary dynamic adjustment strategy is employed, which enables the online prediction model to effectively relearn previously unmodeled samples while shielding external interference. Experimental results obtained from actual testing demonstrates the validity of the secondary dynamic adjustment strategy for online learning, while also providing higher accuracy levels from the prediction model.Impact Statement-Reliable PV power forecasting has become essential in order to ensure the optimal functioning of the power grid. A large number of PV power prediction methods have been developed. However, there are still a number of key challenges. Especially, the prediction effect of forecasting models that have merely undergone offline training are prone to decline due to various changes, which necessitating that the models ultimately need to be retrained. This paper makes significant contributions to solving the existing problems of neural-network-based PV power prediction models. This achievement enhances the forecasting model that can effectively adapt to the dynamic variations of the PV power without the need for retraining.