Ultra-short-term photovoltaic power forecasting based on TCN contrastive encoding and xLSTM

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Time Series Forecasting

Background:

• Intermittency and non-stationarity of photovoltaic (PV) power pose challenges for high-renewable power systems.
• Conventional forecasting methods and standard long short-term memory (LSTM) networks have limitations in handling diurnal non-stationarity and scalar hidden states.

Purpose of the Study:

• To develop an ultra-short-term PV forecasting model, CL-TCN-xLSTM, that overcomes limitations of existing methods.
• To improve the accuracy and reliability of PV power predictions for enhanced grid stability.

Main Methods:

• Proposed CL-TCN-xLSTM model combining contrastive scene encoding with extended LSTM (xLSTM).
• Implemented trend-relative power ratio decomposition to simplify prediction complexity.
• Utilized a temporal convolutional network (TCN) encoder pre-trained via contrastive learning for scene embedding extraction.
• Employed xLSTM with matrix memory to capture multi-timescale fluctuations for forecasting.

Main Results:

• CL-TCN-xLSTM achieved the lowest Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) across forecast horizons from 1 to 4 hours.
• Demonstrated significant reductions in MAE (25.8%) and RMSE (24.8%) compared to the TCR-Reformer baseline.
• Showcased superior error accumulation characteristics and strong cross-site generalization capabilities.
• Effectively handled dawn and dusk transitions without abrupt error spikes, with accurate tracking of power fluctuations on volatile days.

Conclusions:

• CL-TCN-xLSTM offers a robust solution for ultra-short-term PV power forecasting, outperforming existing models.
• The trend-relative power ratio decomposition, xLSTM memory, contrastive pre-training, and TCN encoder are critical components for the model's success.
• The model's ability to capture meaningful weather regimes and generalize across sites highlights its practical applicability in renewable energy integration.