Online Masked Transfer Learning with Cell State Enhanced LSTM: Application in Photovoltaic Power Forecasting and Data Completion

In response to the challenges of integrating photovoltaic (PV) power into the grid, this study introduces an online masked training and transfer learning data completion algorithm (OMTL). This algorithm features a Cell-State-Enhanced-LSTM (CSE-LSTM), which demonstrates high accuracy in long timestep PV power forecasting. Data missing issues are common in practical PV forecasting and can severely impact prediction accuracy, especially when continuous data is missing for days. To address various data missing scenarios, we developed the OMTL-CSE-LSTM. This system utilizes masked training with partial gradient to prevent missing data time points from affecting gradient descent. For extended periods of continuous data missing, it employs transfer learning techniques for data completion. Within OMTL, we adopted a novel LSTM variant that strengthens the cell state containing long-term information, which performs better in long-step PV predictions. In our case study, we compared the accuracy of masked training with pre-prediction data preprocessing methods for missing data. Furthermore, we compared the accuracy of OMTL-CSE-LSTM with other neural network-based OMTL approaches. The experimental results demonstrate that OMTL-CSE-LSTM not only significantly exceeds grid connection standards across various metrics but also surpasses other comparative neural network methods in prediction accuracy at different timesteps.