Key Takeaways & Executive Findings
- •• • Irradiance dominates PV power output with a Pearson coefficient of r = 0.95, while humidity (r = −0.53) and temperature (r = 0.52) provide secondary predictive leverage; wind speed (r = 0.04) and wind direction (r = 0.03) fall below the 0.2 relevance threshold and are excluded, reducing input dimensionality and training convergence time without sacrificing accuracy. • • Against a BiLSTM baseline, the TBiLSTM reduces MAE by 37.16%, MAPE by 55.59%, MSE by 52.11%, and RMSE by 33.75%, while raising R² by 4.03%—a margin that directly translates to lower imbalance penalties in day-ahead market bidding and reduced reserve procurement for grid operators. • • Relative to CNN-BiLSTM, the hybrid architecture still achieves MAE reduction of 12.43%, MAPE reduction of 8.79%, MSE reduction of 20.98%, RMSE reduction of 11.13%, and R² improvement of 2.13%, demonstrating that multi-head self-attention captures cross-time-step dependencies that convolutional kernels cannot resolve under rapid cloud transients. • • The largest single-model gain is against standalone CNN (MAE −47.63%, MAPE −68.80%, MSE −68.47%, RMSE −45.49%), confirming that pure convolutional feature extraction is inadequate for non-stationary PV power series and that recurrent bidirectional modeling remains essential for capturing diurnal ramp dynamics.
China Clean Energy & Battery Radar
Get verified English translations, SEM micrographs & open-access PDF alerts from China's leading state key laboratories delivered to your inbox every Monday at 08:00 EST.
Abstract
Photovoltaic (PV) power forecasting is fundamentally constrained by the stochastic volatility of irradiance, temperature, and atmospheric pressure, which degrades the accuracy of conventional statistical and single-branch recurrent architectures. This study proposes a TBiLSTM hybrid network that fuses multi-temporal features through a Transformer encoder and a bidirectional long short-term memory (BiLSTM) network. Input meteorological sequences—irradiance, humidity, temperature, and pressure—are min-max normalized to [0,1] and augmented with learnable positional encodings to preserve temporal ordering. The Transformer encoder applies multi-head self-attention to capture global cross-time-step dependencies, with residual connections and layer normalization stabilizing gradient propagation. The encoded representation is then passed to a BiLSTM, which extracts forward and backward local temporal dynamics; the final hidden state is mapped to the power sequence via a fully connected layer. Pearson correlation analysis on 3002 samples identifies irradiance (r = 0.95), humidity (r = −0.53), temperature (r = 0.52), and pressure (r = 0.25) as the dominant features, while wind speed (r = 0.04) and wind direction (r = 0.03) are excluded. On two public datasets, the proposed model reduces MAE by 37.16%, 47.63%, and 12.43% versus BiLSTM, CNN, and CNN-BiLSTM, respectively; MAPE by 55.59%, 68.80%, and 8.79%; MSE by 52.11%, 68.47%, and 20.98%; and RMSE by 33.75%, 45.49%, and 11.13%, while improving R² by 4.03%, 4.03%, and 2.13%. These results confirm the model's reliability and superiority for grid dispatch and PV plant operation.
1. Introduction
Grid-scale PV penetration has exposed a persistent operational bottleneck: the stochastic coupling of cloud transients, ambient temperature, and atmospheric pressure produces power ramps that conventional forecasting pipelines cannot track with sufficient fidelity. Statistical mapping methods depend on complete and high-quality historical records and degrade sharply under sensor dropout or anomalous weather; physical models require precise tilt, azimuth, and panel-level meteorological parameters that are rarely available at fleet scale; and single-branch recurrent networks—RNN and LSTM—suffer from information loss over long input horizons or are constrained by layer and unit counts that cap representational capacity. The result is a systematic gap between scheduled and delivered power, which propagates into imbalance charges, reserve activation costs, and reduced renewable curtailment headroom.
The TBiLSTM protocol addresses this bottleneck by decoupling global and local temporal extraction. A Transformer encoder with multi-head self-attention and learnable positional encoding models cross-time-step dependencies in parallel, eliminating the gradient-vanishing constraint of sequential RNN processing. Its output, stabilized by residual connections and layer normalization, is then fed to a BiLSTM that captures forward and backward local dynamics, and a fully connected layer maps the final hidden state to the power sequence. Pearson screening on 3002 samples retains only irradiance, humidity, temperature, and pressure, pruning wind speed and wind direction below the 0.2 relevance threshold. Across two public datasets, this architecture reduces MAE by 12.43–47.63% and RMSE by 11.13–45.49% against CNN-BiLSTM, CNN, and BiLSTM baselines, establishing a measurable accuracy floor for dispatch-grade PV forecasting.
Loading authentic research manuscript (Pages 1–5)...
LI Wanghui, LI Zhendong, LI Shuai, HU Jinchao (2026). TBiLSTM Hybrid Network for Photovoltaic Power Forecasting Based on Multi-Temporal Feature Fusion. Acta Energiae Solaris Sinica. https://doi.org/10.19912/j.0254-0096.tynxb.202608_9682
Research & Educational Purpose Only: The translations, structured abstracts, analytical annotations, and data reports provided by SinoGreenTechare intended exclusively for academic research, internal corporate R&D, and educational benchmarking. They do not constitute formal engineering, chemical safety, legal, or professional advice.
Copyright & Intellectual Property Notice: Original copyright of the underlying source articles and experimental data remains with the respective authors, institutions, and original publishing journals. SinoGreenTech claims intellectual property only over its proprietary translations, analytical syntheses, and AEO structured enhancements in accordance with international fair use and academic citation principles.
Frequently Asked Questions
What is the dominant failure mechanism of standalone BiLSTM under rapid cloud transients, and how does the Transformer encoder mitigate it?
Standalone BiLSTM processes the sequence sequentially, so long input horizons cause gradient vanishing and information loss across distant time steps; it also cannot resolve non-local coupling between irradiance drops and delayed power ramps. The Transformer encoder applies multi-head self-attention in parallel across the entire input, assigning weights to all time steps simultaneously, with residual connections and layer normalization stabilizing gradients. This yields a 37.16% MAE reduction and 33.75% RMSE reduction versus BiLSTM, directly improving ramp-event tracking.
Why were wind speed and wind direction excluded despite being standard meteorological inputs?
Pearson correlation analysis on 3002 samples returned r = 0.04 for wind speed and r = 0.03 for wind direction, both below the 0.2 threshold for weak or no correlation. Retaining them would increase input dimensionality, slow convergence, and risk overfitting without predictive gain. The retained features—irradiance (r = 0.95), humidity (r = −0.53), temperature (r = 0.52), and pressure (r = 0.25)—carry the measurable signal.
Does the TBiLSTM architecture maintain a cost-accuracy advantage over CNN-BiLSTM in operational deployment?
Against CNN-BiLSTM, TBiLSTM reduces MAE by 12.43%, MAPE by 8.79%, MSE by 20.98%, RMSE by 11.13%, and improves R² by 2.13%. The incremental computational cost is confined to the self-attention block, which parallelizes across time steps and avoids the sequential penalty of pure recurrent stacks. For a utility-scale plant, an 8.79% MAPE reduction directly lowers imbalance settlement and reserve activation charges, offsetting the added inference cost.
What is the scalability bottleneck when extending this model from two public datasets to a multi-site PV fleet?
The model requires consistent multi-dimensional inputs—irradiance, humidity, temperature, and pressure—at uniform sampling intervals. Cross-site heterogeneity in sensor calibration, tilt angle, and microclimate shifts the input distribution, which can degrade the learned attention weights. Retraining or fine-tuning per site is necessary; the 3002-sample correlation analysis must be repeated to confirm that the 0.2 relevance threshold still excludes wind features at each location. Without this, the reported 12.43–47.63% MAE gains are not transferable.
How does min-max normalization to [0,1] affect model stability and the interpretability of the reported error metrics?
Min-max normalization prevents high-magnitude features such as pressure from dominating gradient updates, which would otherwise suppress learning on lower-magnitude but higher-relevance features like humidity. The inverse transform in Eq. (3) restores predictions to original power units before MAE, MAPE, MSE, and RMSE computation, so all reported reductions—37.16% MAE, 55.59% MAPE, 52.11% MSE, 33.75% RMSE versus BiLSTM—are in physical units and directly comparable to baseline performance.
Related Chinese Research & Cross-Citations
Wind Turbine Gearbox Fault Diagnosis Method Based on Improved CNN-XGBoost Fusion Model Under Gramian Angular Difference Field
Gearbox failures account for 20–30% of total wind turbine faults and incur maintenance costs equivalent to 10–15% of overall turbine value. Conventional vibration diagnostic pipelines—complementary ensemble empirical mode decomposition with singular value energy spectrum, time-varying filtering empirical mode decomposition, and Teager energy spectrum analysis—remain bounded below 90% accuracy and depend on expert-driven feature engineering that is sensitive to non-stationary operating conditions and noise. This study proposes an intelligent diagnostic architecture that converts one-dimensional gearbox vibration signals into two-dimensional images via Gramian angular difference field (GADF) transformation, preserving intrinsic temporal correlation and time-frequency structure while exploiting matrix sparsity to suppress interference. An improved convolutional neural network (CNN) extracts multi-dimensional features: a convolutional block attention module (CBAM) is embedded in the convolutional layers to weight critical channels and focus on fault-sensitive spatial regions, and a modified βc-ACONC activation function replaces ReLU to mitigate neuron necrosis and enable selective activation. The extracted composite features are then fed into an XGBoost network whose hyperparameters are optimized by an improved sparrow search algorithm (ISSA). Validation on a laboratory wind turbine gearbox dataset yields diagnostic accuracy exceeding 99%, demonstrating robust fault identification capability under complex operating conditions.
Joint Forecasting of Wind and Photovoltaic Power Considering Complementarity
The inherent spatiotemporal complementarity between wind and solar resources offers a theoretical basis for improving renewable power forecasting accuracy. This study proposes a joint wind-photovoltaic (PV) power forecasting strategy that explicitly exploits this complementarity. A bidirectional long short-term memory (BiLSTM) neural network serves as the baseline forecasting model, and a novel sorting and comparative optimization (SCO) algorithm is developed to optimize the model's hyperparameters. The SCO algorithm ranks individuals in ascending order and compares adjacent fitness values to escape local optima, a known deficiency in conventional metaheuristics such as genetic algorithms and particle swarm optimization. For wind farms and PV plants exhibiting significant complementarity, the joint forecasting strategy first aggregates their power outputs, normalizes the combined signal, and then feeds it into the optimized BiLSTM model. Experimental results demonstrate that the proposed SCO-BiLSTM model reduces the eMAPE by 10.313% compared with PSO-BiLSTM. Furthermore, joint forecasting under SCO-BiLSTM lowers the eRMSE by 27.443% relative to standalone PV power forecasting. The study also establishes that forecasting accuracy improves with stronger wind-solar complementarity but degrades as the forecasting horizon extends. These findings confirm that exploiting complementarity in joint forecasting substantially enhances predictive performance for renewable energy integration.
Ultra-Short-Term Wind Power Forecasting Based on Fluctuation Continuation Scenario Identification
Existing ultra-short-term wind power forecasting methods exhibit limited performance due to insufficient extraction of fluctuation information and inadequate analysis of evolution patterns. This paper proposes an ultra-short-term wind power forecasting method based on fluctuation continuation scenario identification. First, the coupling mechanism of wind power fluctuations under multiple turbulence processes is investigated, and historical wind power dynamics are decoupled into a combination of nonlinear and linear fluctuation components. A fluctuation continuation concept is introduced, and the future continuation scale of wind power fluctuations is derived from nonlinear and linear decoupling parameters, thereby classifying fluctuation continuation scenarios. A sparse neural network (SNN) oriented to high-dimensional sparse features is constructed to identify historical fluctuation continuation scenarios, and ultra-short-term power forecasting is conducted separately for each scenario. Validation using measured wind speed and power data from three wind farms shows that, compared with baseline models, the proposed model improves root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) by at least 1.46%, 2.44%, and 14.67%, respectively, demonstrating superior accuracy and stability. The method addresses the limitations of signal decomposition techniques that lack physical interpretability and are sensitive to hyperparameters, and overcomes the high-dimensional sparsity challenges faced by traditional scenario identification models.
Multi-Classifier Open Adversarial Network for Rolling Bearing Fault Diagnosis in Wind Turbine Generator Systems
Rolling bearings in wind turbine generator systems operate under variable speed and load conditions that induce significant data distribution shifts between training and field data, while unknown fault modes absent from the source domain are frequently misclassified as known classes. This study proposes a multi-classifier open adversarial network (MCOAN) for open-set fault diagnosis. Within an adversarial domain adaptation framework, a K-way classifier and an additional K+1-way one-vs-all classifier independently estimate target-sample similarity to the source domain. These similarity scores drive a dynamic weighting mechanism that adaptively reweights target samples during open-set adversarial training and supplies per-sample dynamic thresholds for known/unknown discrimination, thereby promoting cross-domain alignment of shared-class features while suppressing negative transfer from unknown samples. A non-adversarial domain classifier is introduced to stabilize dynamic weight estimation. Validation on two datasets demonstrates high-precision shared-class distribution alignment and unknown-class recognition with favorable robustness. The method removes reliance on empirically preset thresholds that plague conventional open-set back-propagation approaches, where a fixed threshold of 0.5 in the binary cross-entropy adversarial loss provides no per-sample adaptivity. By coupling K-way and K+1-way similarity estimates, MCOAN achieves simultaneous known-class alignment and unknown-class separation without prior knowledge of the unknown-class cardinality, addressing a persistent bottleneck in wind turbine drivetrain condition monitoring where unanticipated bearing failure modes emerge under field conditions.
Unsupervised Automated Identification Method for Abnormal States of Wind Turbine Gearboxes
Addressing the scarcity of labeled data for training classification models in wind turbine planetary gearbox anomaly identification, this study proposes an unsupervised automated detection method. Log Mel-band energy features are extracted from raw vibration signals and fed into an unsupervised anomaly recognition model centered on a U-net autoencoder. A health-state threshold is established based on reconstruction error between model input and output, enabling anomaly identification. The method is validated using factory gearbox test data and operational data from a wind farm in Yangtouya, Shanxi. For factory gearboxes, dual validation is performed using a spectrum amplitude modulation-based signal processing method. Results demonstrate that the proposed method achieves 93.34% recognition accuracy on both factory and wind farm test sets, confirming its capability to automatically and correctly separate abnormal wind turbine gearboxes. The approach eliminates reliance on labeled fault data, offering a scalable solution for full-lifecycle health monitoring, from factory acceptance testing to in-service early anomaly detection, adaptable across different operating conditions and turbine models.
Improved Adaptive Super-Twisting Sliding Mode Control for Permanent Magnet Synchronous Motors
Adaptive super-twisting sliding mode control (ASTSMC) for permanent magnet synchronous motors (PMSM) suffers from prolonged convergence and insufficient disturbance rejection under complex operating conditions. This paper proposes an improved ASTSMC incorporating a fixed-time disturbance observer (FTDO). A power term is introduced into the adaptive super-twisting controller to accelerate convergence far from the origin, while the discontinuous sign function is replaced by a continuous h(s) function to mitigate chattering. The FTDO ensures disturbance estimation converges within a fixed time independent of initial states, overcoming the limitations of traditional and finite-time observers. The estimated disturbance is fed forward to the sliding mode controller for compensation, enhancing robustness. The FTDO design is based on an auxiliary state variable z and its derivative, with error dynamics analyzed via Lyapunov stability. Comparative simulations against conventional disturbance observers and sliding mode controllers validate the proposed strategy. The results demonstrate shorter convergence time, improved dynamic performance, and superior disturbance rejection, making the approach suitable for high-performance servo systems. The method addresses the critical need for robust, fast-response control in electric vehicles, rail transit, and aerospace applications where PMSM drives face significant uncertainties and external disturbances.