LFTL: Lightweight feature transfer learning with channel-independent LSTM for distributed PV forecasting

Abstract

Distributed photovoltaic (PV) power forecasting in newly installed systems faces challenges due to inherent stochastic volatilities and limited historical data. This paper proposes a lightweight feature transfer learning (LFTL) method that enables rapid and accurate forecasting of new distributed PVs. Firstly, the raw fluctuating PV data are preprocessed through decomposition to separate low- and high-frequency components. These components are then multi-scale segmented to capture diverse temporal characteristics. Following feature compression and LSTM temporal modeling, the informative features from the source domain enable lightweight transfer. For the target domain, a channel-independent encoder is designed to prevent negative interactions between heterogeneous frequencies. The frequency-fused segment-independent decoder equipped with positional embeddings enables local temporal analysis and reduces error accumulation of multi-step forecasts. LFTL trains with a joint training strategy to avoid negative transfer caused by domain disparity. LFTL consistently outperforms state-of-the-art time-series forecast models while maintaining a relatively low computational overhead based on real-world distributed PV data.