CT-PatchTST: Channel-Time Patch Time-Series Transformer for Long-Term Renewable Energy Forecasting

TL;DR

CT-PatchTST addresses the challenge of long-horizon renewable energy forecasting by jointly modeling inter-channel correlations and temporal dynamics in multivariate time series. It extends PatchTST with a dual-channel/time attention encoder, RevIN normalization, and patch-based input representations, yielding more accurate wind and solar forecasts. Evaluations on real-world Danish offshore/onshore wind and solar data show consistent improvements over PatchTST and other baselines, underscoring robustness across patch lengths and forecast horizons. The results imply improved predictive capabilities for energy storage coordination and grid operation, enabling proactive ESS deployment and reduced dispatch risk in high-renewable scenarios.

Abstract

Accurate forecasting of renewable energy generation is fundamental to enhancing the dynamic performance of modern power grids, especially under high renewable penetration. This paper presents Channel-Time Patch Time-Series Transformer (CT-PatchTST), a novel deep learning model designed to provide long-term, high-fidelity forecasts of wind and solar power. Unlike conventional time-series models, CT-PatchTST captures both temporal dependencies and inter-channel correlations-features that are critical for effective energy storage planning, control, and dispatch. Reliable forecasting enables proactive deployment of energy storage systems (ESSs), helping to mitigate uncertainties in renewable output, reduce system response time, and optimize storage operation based on location-specific flow and voltage conditions. Evaluated on real-world datasets from Denmark's offshore wind, onshore wind, and solar generation, CT-PatchTST outperforms existing methods in both accuracy and robustness. By enabling predictive, data-driven coordination of ESSs across integrated source-grid-load-storage systems, this work contributes to the design of more stable, responsive, and cost-efficient power networks.

Figures (6)

• Figure 1: Overview of CT-PatchTST model. After the multivariate time series is processed through channel attention, the inter-channel relationships are learned, resulting in the generation of a new multivariate time series. This transformed series is then passed through time attention, where each channel is treated independently to capture temporal dependencies. Finally, the prediction results are produced based on the integrated outputs of the channel and time attention mechanisms.
• Figure 2: Schematic architecture of the CT-PatchTST. After the multivariate time series is processed through channel attention, inter-channel relationships are learned, producing a new multivariate series. This transformed series is then passed through time attention, where temporal dependencies are captured independently for each channel. The final prediction results are obtained by integrating the outputs of both attention mechanisms.
• Figure 3: The schematic of patch and projection. Using patch to expand the input semantic information, and using projection to expand the input dimension to improve the fitting ability of the model
• Figure 4: The forecasting performance (MSE) is evaluated on the dataset using CT-PatchTST and PatchTST models, with a fixed look-back window of 336 and forecasting lengths $h \in \{96, 192, 336, 720\}$. The patch length $P$ is varied across $\{4, 8, 12, 16, 24, 32, 40\}$ to investigate its impact on model performance.
• Figure 5: Attention maps of the Channel Attention mechanism and the forecasting results of selected time series, generated using CA-PatchTST. Experiments are conducted with a look-back window of 336 and a forecasting length of 96. Additionally, the average channel attention across all patches is computed to provide a comprehensive view of the attention distribution.