A Hybrid Model for Short-Term Photovoltaic Power Forecasting Based on CNN-BiLSTM and Scenario-Adaptive LSTM

View at publisher Summarise with VIRAT
Authors: Junjie Wu
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper proposes a prediction model based on time series features, focusing on multi-model architecture design and feature fusion methods. First, a theoretical model is constructed to analyze short- and long-term periodic features and to clarify the deviation mechanism between theoretical and actual values. Next, a CNN-BiLSTM model based on sliding windows and time information encoding is designed. A 1D convolutional layer is used to extract local features, and a bidirectional long short-term memory network captures the temporal dependencies. The subsequent period predictions are achieved through global average pooling and a fully connected layer. The training process is optimized using a mean squared error loss function and cosine annealing strategy. Finally, real-time feature data is incorporated, and scenarios are classified using principal component analysis for dimensionality reduction and clustering algorithms. An improved long short-term memory network model is built, dynamically adjusting feature weights to enhance prediction accuracy under complex conditions. The model effectively learns temporal patterns without relying on external data, and after feature fusion, further reduces prediction errors. It accurately captures both periodicity and volatility, demonstrating strong generalization ability and prediction stability across multiple scenarios.

Fields of Study & Subjects

Solar Radiation and Photovoltaics Energy Load and Power Forecasting Stock Market Forecasting Methods

Publication details

Journal / source
IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
Pages
965–971
Publication date
28 January 2026
Language
EN