Open Access

ARTICLE

Research on Ultra-Short-Term Photovoltaic Power Forecasting Based on Parallel Architecture TCN-BiLSTM with Temporal-Spatial Attention Mechanism

Hongbo Sun¹, Xingyu Jiang^1,*, Wenyao Sun¹, Yi Zhao¹, Jifeng Cheng², Xiaoyi Qian¹, Guo Wang³

1 School of Electrical Engineering, Shenyang Institute of Engineering, Shenyang, 110136, China
2 State Grid Liaoning Electric Power Co., Ltd., Electric Power Research Institute, Shenyang, 110006, China
3 School of Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou, 730070, China

* Corresponding Author: Xingyu Jiang. Email:

(This article belongs to the Special Issue: Advances in Renewable Energy Systems: Integrating Machine Learning for Enhanced Efficiency and Optimization)

Energy Engineering 2026, 123(4), 13 https://doi.org/10.32604/ee.2025.073012

Received 09 September 2025; Accepted 06 November 2025; Issue published 27 March 2026

Abstract

The accuracy of photovoltaic (PV) power prediction is significantly influenced by meteorological and environmental factors. To enhance ultra-short-term forecasting precision, this paper proposes an interpretable feedback prediction method based on a parallel dual-stream Temporal Convolutional Network-Bidirectional Long Short-Term Memory (TCN-BiLSTM) architecture incorporating a spatiotemporal attention mechanism. Firstly, during data preprocessing, the optimal historical time window is determined through autocorrelation analysis while highly correlated features are selected as model inputs using Pearson correlation coefficients. Subsequently, a parallel dual-stream TCN-BiLSTM model is constructed where the TCN branch extracts localized transient features and the BiLSTM branch captures long-term periodic patterns, with spatiotemporal attention dynamically weighting spatiotemporal dependencies. Finally, Shapley Additive explanations (SHAP) additive analysis quantifies feature contribution rates and provides optimization feedback to the model. Validation using operational data from a PV power station in Northeast China demonstrates that compared to conventional deep learning models, the proposed method achieves a 17.6% reduction in root mean square error (RMSE), a 5.4% decrease in training time consumption, and a 4.78% improvement in continuous ranked probability score (CRPS), exhibiting significant advantages in both prediction accuracy and generalization capability. This approach enhances the application effectiveness of ultra-short-term PV power forecasting while simultaneously improving prediction accuracy and computational efficiency.

---

Keywords

---