A missing-data-robust hybrid neural network for photovoltaic power forecasting using improved Golden Jackal optimization

Technical Summary: A Missing-Data-Robust Hybrid Neural Network for Photovoltaic Power Forecasting

Problem Statement
Accurate short-term photovoltaic (PV) power forecasting is critical for power system scheduling and renewable energy integration. However, the reliability of forecasting models is frequently compromised by missing values in monitoring data. These gaps arise from communication interruptions, sensor faults, and recording errors. If missing data are not properly repaired prior to training, models may learn distorted temporal patterns, leading to reduced accuracy and robustness. While existing studies have focused on model architecture improvements or parameter optimization, there is a need for a comprehensive framework that addresses data imputation, hybrid deep learning feature extraction, and intelligent hyperparameter tuning simultaneously.

Methodology
The proposed framework follows a three-stage pipeline: data preprocessing, hybrid model construction, and intelligent optimization.

1. Data Preprocessing and Imputation:

   • Variable Selection: Pearson correlation analysis was performed on meteorological and operational variables (e.g., irradiance, temperature, wind speed) to screen for features with strong correlations to PV power, eliminating redundant inputs.
   • Imputation Strategy: Four imputation methods were systematically evaluated under varying missing-data rates (10%, 20%, and 30%): Linear Interpolation, K-Nearest Neighbor (KNN), Multiple Imputation, and Random Forest Imputation.
   • Outlier Handling: A 3-sigma criterion was applied to detect and correct outliers, replacing anomalous points with previous valid observations to maintain temporal continuity.

2. Hybrid Forecasting Model (TCN-BiGRU-Attention):

   • Temporal Convolutional Network (TCN): Utilized to extract local temporal features and long-term dependencies through causal and dilated convolutions, avoiding gradient vanishing issues common in RNNs.
   • Bidirectional Gated Recurrent Unit (BiGRU): Employed to capture bidirectional sequence dependencies, considering both past and future information relative to the current time step.
   • Attention Mechanism: Integrated to assign adaptive weights to key time steps, allowing the model to focus on critical features (e.g., specific weather conditions) while suppressing noise.
   • Output: A fully connected layer generates the final regression prediction.

3. Optimization via Improved Golden Jackal Optimizer (IGJO):

   • The standard Golden Jackal Optimizer (GJO) was enhanced with six strategies to improve global exploration, local exploitation, and convergence speed:
     • Adaptive energy coefficient (nonlinear decreasing update).
     • Enhanced Lévy Flight for broader search space jumps.
     • Hybrid learning strategy (dynamic selection of update rules).
     • Adaptive weight update for balancing exploration and exploitation.
     • Local search mechanism with boundary constraints.
     • Chaos mapping-guided exploration to avoid local optima.
   • IGJO was used to automatically tune hyperparameters, including learning rate, BiGRU neuron count, attention dimension, and regularization coefficients.

Key Contributions

• Imputation Evaluation: A comparative analysis demonstrating that Multiple Imputation offers the most robust performance across varying missing-data rates, particularly in high-missing scenarios, while Random Forest provides strong nonlinear modeling capabilities.
• Architectural Innovation: The construction of a TCN-BiGRU-Attention hybrid network that effectively fuses local temporal patterns, bidirectional dependencies, and key feature weighting.
• Algorithmic Improvement: The development of the IGJO algorithm, which outperforms the standard GJO in convergence speed and optimization accuracy for neural network hyperparameter tuning.
• Seasonal Robustness: Validation of the framework across four distinct seasons, demonstrating consistent performance improvements under varying climatic conditions (e.g., high irradiance in summer vs. low irradiance in winter).

Experimental Results
Experiments were conducted using 15-minute interval data from a grid-connected PV station in Ningxia, China, covering the year 2019.

• Imputation Performance: Multiple Imputation achieved the lowest Mean Squared Error (MSE) across all missing rates (e.g., MSE ~0.0002 at 10% missing), significantly outperforming Linear Interpolation and KNN, especially as missing rates increased to 30%.
• Seasonal Forecasting: The IGJO-optimized model consistently outperformed both the baseline network and the standard GJO-optimized model across all seasons.
  • Spring: MAE reduced by 55.11% compared to the baseline.
  • Summer: MAE reduced by 78.85% compared to the baseline; MSE reduced by 19.47% compared to GJO.
  • Autumn: The most significant improvement was observed here, with MAE reduced to 24.64% of the baseline value (a 75.36% reduction) and RMSE reduced by 67.32%.
  • Winter: MAE reduced by 48.19% compared to the baseline.
• Model Comparison: Against single models (TCN, BiGRU, CNN, LSTM) and hybrid baselines (CNN-LSTM, TCN-GRU), the IGJO-TCN-BiGRU-Attention model achieved the lowest error metrics. Specifically, its MAE and RMSE were 72.8% and 12.8% lower, respectively, than the suboptimal model (LSTM).

Significance and Claims
The paper claims that the proposed framework provides a robust solution for PV power forecasting in the presence of incomplete monitoring data. By integrating multiple imputation for data restoration, a hybrid deep learning architecture for feature extraction, and an improved meta-heuristic optimizer for parameter tuning, the system achieves higher accuracy and stability than existing methods. The authors assert that this approach offers technical support for power system dispatch and decision-making when data quality is compromised. The study concludes that Multiple Imputation and Random Forest are the preferred strategies for data restoration, and the IGJO-optimized hybrid model is highly adaptable to seasonal variations, making it suitable for practical engineering applications in renewable energy management. Future work is suggested to extend the method to multi-station datasets and online deployment.