De-Trend First, Attend Next: A Mid-Term PV forecasting system with attention mechanism and encoder–decoder structure

Abstract

Performing accurate predictions on photovoltaic power generation is a crucial factor in complementary power generation scheduling. How to resolve the problem of extracting hidden features and correcting abnormal data is the key factor of improving mid-term prediction accuracy. This study proposes a mid-term PV forecasting system using the De-Trend First, Attend Next strategy. The prediction system employs the detrending before attending strategy. It first adopts and corrects abnormal time series data, and then reconstructs the corrected time series data into trend and seasonal components. After data is constructed, different models are applied to trend and seasonal data for separately predictions, and such predictions then be reconstructed into a realistic prediction result. To find hidden features and seasonal trends of seasonal components, we propose a new model. This model is constructed with an encoder–decoder structure temporal convolution, and attention mechanism. This study evaluates the prediction system using data from a photovoltaic power station in Australia. The experimental results show that the proposed model has a Coefficient of Determination value of 0.992, which represents a 73% improvement in the mean squared error index compared to the baseline model. In summary, the experiment result demonstrates that the system has a good capability of providing and predicting accurate data, which plays a significant role in power grid dispatch.