Pattern Sequence Neural Network For Solar Power Forecasting

We propose a new approach for time series forecasting, called PSNN, which combines pattern sequences with neural networks. It is a general approach that can be used with different pattern sequence extraction algorithms. The main idea is to build a separate prediction model for each pattern sequence type. PSNN is applicable to multiple related time series. We demonstrate its effectiveness for predicting the solar power output for the next day using Australian data from three data sources - solar power, weather and weather forecast. In our case study, we show three instantiations of PSNN by employing the pattern sequence extraction algorithms PSF, PSF1 and PSF2. The results show that PSNN achieved the most accurate results.