This study introduces a Cross-Variable Attention mechanism to explicitly reconstruct nonlinear photothermal couplings via dynamic attention weights and provides an accurate, generalizable, and physically interpretable solution for collaborative multi-station distributed PV energy forecasting.
Abstract
Accurate photovoltaic (PV) energy forecasting is vital for grid stability and the global low-carbon transition. However, existing data-driven and channel-independent PV energy forecasting models struggle to capture nonlinear meteorological couplings, heterogeneous physical scales across stations, and high-frequency non-stationary fluctuations. To address these limitations, this study proposes a Physics-Guided Cross-Variable Temporal Transformer architecture. Building upon a channel-independent foundation, we introduce a Cross-Variable Attention mechanism to explicitly reconstruct nonlinear photothermal couplings via dynamic attention weights. To resolve multi-station physical scale discrepancies, a Physical Feature-wise Linear Modulation network utilizes installed capacity as a static prior for adaptive cross-station scale alignment. During optimization, a Time Dynamics-Aware Perceiving Loss jointly penalizes absolute errors and first-order time differences, constraining the network’s tracking ability for transient ramping. Experiments demonstrate that the proposed architecture overcomes traditional channel-isolation limitations. The model achieves a 21.6% reduction in MSE compared to PatchTST, a 44.0% reduction compared to Autoformer, and a 2.7% improvement in R2 over Informer. This provides an accurate, generalizable, and physically interpretable solution for collaborative multi-station distributed PV energy forecasting.
Accurate wind power forecasting is essential for renewable-energy accommodation, low-carbon dispatch, and the sustainable operation of modern power systems. However, wind power series exhibit pronounced non-stationarity, strong volatility, and multi-scale evolution, making long-term trends and short-term disturbances difficult to characterize jointly. In addition, multi-step forecasting errors tend to accumulate with increasing horizons, degrading model accuracy and stability. To address these issues, this study proposes CoFFormer, a collaborative frequency-domain-enhanced network for non-stationary wind power forecasting. The model reduces input modeling complexity, strengthens collaborative representation of heterogeneous temporal information, and suppresses output-stage error accumulation. Specifically, embedded series decomposition mitigates coupling interference between trend and fluctuation components. Differentiated temporal modeling and dynamic gating then adaptively coordinate the contributions of different feature representations, while frequency-domain residual compensation enhances the recovery of periodic structures and local oscillations. Experiments on ETTh2, wind_speed, WindPower, and Location2 demonstrate strong competitiveness across forecasting horizons. CoFFormer achieves MSE/MAE values of 0.0957/0.2238 and 0.1508/0.2889 on ETTh2 for 12- and 24-step forecasting, and 0.0617/0.1490 and 0.3838/0.3948 on WindPower for 3- and 24-step forecasting, outperforming most baselines. Ablation studies confirm the effectiveness and synergy of each component, providing an effective solution for high-accuracy multi-step forecasting of complex non-stationary wind power series.
The rapid expansion of photovoltaic (PV) systems poses significant challenges to grid stability. Hybrid Energy Systems (HES) are intended to alleviate this volatility, yet their coordinated dispatch often remains suboptimal due to communication delays and ramp-rate constraints. Accurate ultra-short-term PV power forecasting is therefore essential, as it enables preemptive control and timely dispatch adjustments that unlock the full potential of HES. In this study, we propose a novel AI hybrid forecasting framework that integrates a rule-based model with a Decomposition Linear (DLinear) Long Short-Term Memory (LSTM) deep learning core, representing, to the best of our knowledge, a novel integration of a decomposition-based linear model (DLinear) with LSTM networks for ultra-short-term PV power forecasting. The DLinear component decomposes the time series into trend and remainder sequences, which are then independently modeled by separate LSTM networks to capture distinct dynamics. Using data from a 300 kWp PV power station, the framework achieves an average daily prediction accuracy exceeding 93% for both 5-min and 15-min horizons. The model reliably tracks power variations under sunny and rainy conditions, while under volatile cloudy weather its accuracy decreases but still captures essential fluctuation patterns. These results demonstrate the potential of the proposed framework for improving the dispatch and operational reliability of hybrid energy systems. However, further validation across additional seasons and sites is needed to establish broader generalizability.
Fu-Yan Huang, Gang Xiao, Keqin Wang et al.· Energies· 0 citations
Photovoltaic (PV) power generation is affected by meteorological and geographical factors, displaying volatility, nonlinearity, and characteristics across multiple timescales. Accurate short-term forecasting of PV power is essential for enhancing PV resource utilization in the power grid and maintaining the quality of renewable energy integration. This paper introduces a Multi-Variable Time-Frequency Synergistic Fusion network model, enhanced by the Triangular Wave Zebra Algorithm (TWZA), for short-term PV power forecasting. To tackle the challenges of disjointed time-frequency feature fusion and limited generalization in complex fluctuation scenarios, a dual-parallel branch structure is developed to synergistically model temporal and spectral features in both time and frequency domains. In the time-domain branch, a kLSTM encoder is employed, which integrates block embedding and a Multi-Variable Correlation Attention mechanism to uniformly model variable coupling relationships and long-term dependencies. The frequency-domain branch incorporates the frequency decomposition architecture of the Multi-order Kolmogorov–Arnold Network (KAN), decomposing the power signal into multi-scale frequency components. By combining multi-order KAN representation learning with deep separable convolutions, it extracts nonlinear spectral features. A cross-domain attention-based time-frequency fusion mechanism is developed for adaptive and collaborative integration of time-domain and frequency-domain features. Second, to address the inability of time-frequency models to adaptively tune hyperparameters, we introduce TWZA to adaptively optimize time-frequency fusion weights, network hyperparameters, and convolution kernel sizes. This enhances the model's global search and local exploration capabilities while reducing training loss. Extensive experimental validation demonstrates that the proposed model outperforms all baseline models in terms of R2, root mean square error, and mean absolute error, achieving improvements of at least 18.2%, 27.3%, and 14.9%, respectively.
Wan-Nian Wei, Zhiwen Wang, Haoxu Wang et al.· Journal of Renewable and Sus...· 0 citations
Accurate short-term photovoltaic (PV) power forecasting is essential for secure grid operation, economic dispatch, and efficient utilization of renewable energy. However, PV output exhibits strong nonlinearity and nonstationarity due to rapidly varying meteorological conditions, cloud movements, and system uncertainties. Conventional statistical models and single-source deep learning approaches often fail to fully exploit the rich multi-source information available in modern PV plants, such as historical power, on-site meteorological measurements, and clear-sky or numerical weather prediction (NWP) features. In this paper, we propose MSF-TransPV, a Multi-Source Fusion Transformer framework for short-term PV power forecasting. The model adopts a multi-branch temporal encoder that separately processes historical PV output, meteorological variables, and optional clear-sky/NWP-derived features, mapping them into a shared latent space. A cross-variable multi-head attention module is then introduced to explicitly capture the dependencies between PV dynamics and atmospheric conditions, enabling fine-grained interaction across different feature sources. On top of the fused representation, a temporal Transformer encoder models long-range temporal dependencies and feeds a lightweight decoder that supports both point forecasting and probabilistic forecasting via quantile regression. Experimental results on real-world PV datasets demonstrate that MSF-TransPV consistently outperforms persistence, statistical baselines, recurrent neural networks, and vanilla Transformer models in terms of RMSE, MAE, and normalized error metrics, while also providing reliable prediction intervals. These results indicate that explicit multi-source fusion and cross-variable attention are highly effective for PV forecasting under complex and rapidly changing weather conditions.
Xiao-Mei Wang, Pei-Xuan Xu, Xiao-Hui Wang· European Conference on Elect...· 0 citations
Accurate short-term photovoltaic (PV) power forecasting is critical for secure grid operation and economic dispatch, yet its performance is often degraded by non-stationary irradiance fluctuations induced by cloud transients and weather regime shifts. To address this challenge, this paper proposes a physics-guided temporal fusion architecture built on the Temporal Fusion Transformer (TFT) for multi-horizon probabilistic PV forecasting. The model fuses historical plant measurements, Numerical Weather Prediction (NWP) variables, and solar-geometry features, and introduces a cross-attention fusion module to align future meteorological drivers with the most relevant historical context for improved ramp responsiveness. Physical plausibility is promoted by soft physics-consistency regularization, including non-negativity, rated-capacity bounds, clear-sky envelope constraints, and ramp-rate penalties, while uncertainty is quantified via multi-quantile regression to produce calibrated prediction intervals. Experiments on real-world PV datasets across different seasons and weather conditions show that the proposed approach consistently improves deterministic accuracy (MAE and RMSE) and probabilistic performance (PICP and CRPS) over representative baselines, with particularly robust behavior under regime-shift and ramp events. Overall, combining attention-based temporal fusion with physics-guided learning provides a practical and scalable solution for reliable probabilistic PV power forecasting in highly variable atmospheric conditions.
Pei-Xiang Wu· European Conference on Elect...· 0 citations
Accurate high-resolution photovoltaic power forecasting remains challenging because rapid weather changes and heterogeneous plant configurations produce nonlinear variations and extreme errors. Existing data-driven methods capture complex meteorological relationships but may generate physically inconsistent forecasts, whereas simplified physical models cannot fully represent site specific behavior. This study developed a physics-informed temporal fusion transformer framework combining a transformer–bidirectional long short-term memory forecasting branch with soft irradiance–power and temperature–power constraints. The framework was evaluated using 70,176 observations from Solar Station Site 5, recorded at 15-minute intervals, and independently trained and calibrated across eight solar stations. On the Site 5 test set, it achieved a mean absolute error of 1.4253 megawatts, a root mean squared error of 3.8612 megawatts, and a coefficient of determination of 0.9734. It outperformed standalone temporal fusion transformer, gated recurrent unit, extreme gradient boosting, and random forest; relative to standalone temporal fusion transformer, mean absolute and root mean squared errors decreased by 29.7% and 15.1%, respectively. Sensitivity analysis selected 0.05 as the best nonzero physics weight for Site 5 and revealed a trade-off between predictive accuracy and physical consistency. Residual analysis identified transient and measurement-related outliers. Across the eight stations, coefficients of determination exceeded 0.84 at six sites and reached 0.9804 at Site 6, supporting generalizability across heterogeneous installations following station-specific calibration. Deployment latency remained 0.0864 milliseconds per sample, supporting operational use for renewable-energy integration, grid scheduling, and low-carbon power management.
M. Nasir, Muhammad Farhan Hanif, Muhammad Tahir Hassan et al.· Clean Energy· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
