Robust autoregressive bidirectional gated recurrent units model for short-term power forecasting

Engineering Applications of Artificial Intelligence, Journal Year: 2024, Volume and Issue: 138, P. 109453 - 109453

Published: Oct. 20, 2024

Language: Английский

---

An Informer-based multi-scale model that fuses memory factors and wavelet denoising for tidal prediction

Electronic Research Archive, Journal Year: 2025, Volume and Issue: 33(2), P. 697 - 724

Published: Jan. 1, 2025

Language: Английский

---

Forecasting very short-term power load with hybrid interpretable deep models

Systems Science & Control Engineering, Journal Year: 2025, Volume and Issue: 13(1)

Published: April 1, 2025

Language: Английский

---

Multi-Granularity Autoformer for long-term deterministic and probabilistic power load forecasting

Neural Networks, Journal Year: 2025, Volume and Issue: 188, P. 107493 - 107493

Published: April 24, 2025

Long-term power load forecasting is critical for system planning but constrained by intricate temporal patterns. Transformer-based models emphasize modeling long- and short-term dependencies yet encounter limitations from complexity parameter overhead. This paper introduces a novel Multi-Granularity Autoformer (MG-Autoformer) long-term forecasting. The model leverages Auto-Correlation Attention Mechanism (MG-ACAM) to effectively capture fine-grained coarse-grained dependencies, enabling accurate of fluctuations trends. To enhance efficiency, shared query-key (Q-K) mechanism utilized identify key patterns across multiple resolutions reduce complexity. address uncertainty in forecasting, the incorporates quantile loss function, probabilistic predictions while quantifying uncertainty. Extensive experiments on benchmark datasets Portugal, Australia, America, ISO New England demonstrate superior performance proposed MG-Autoformer point tasks.

Language: Английский

---

A hybrid power load forecasting model using BiStacking and TCN-GRU

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(4), P. e0321529 - e0321529

Published: April 28, 2025

Accurate power load forecasting helps reduce energy waste and improve grid stability. This paper proposes a hybrid model, BiStacking+TCN-GRU, which leverages both ensemble learning deep techniques. The model first applies the Pearson correlation coefficient (PCC) to select features highly correlated with load. Then, BiStacking is used for preliminary predictions, followed by temporal convolutional network (TCN) enhanced gated recurrent unit (GRU) produce final predictions. experimental validation based on Panama’s 2020 electricity data demonstrated effectiveness of achieving an RMSE 29.1213 MAE 22.5206, respectively, R² 0.9719. These results highlight model’s superior performance in short-term forecasting, demonstrating its strong practical applicability theoretical contributions.

Language: Английский

---

A two-stage ultra short-term load forecasting model for comprehensive energy based on FCM

Journal of Physics Conference Series, Journal Year: 2025, Volume and Issue: 3011(1), P. 012002 - 012002

Published: May 1, 2025

Abstract The dynamic scheduling requirements of a two-stage integrated energy system with single-layer variable step size are complex and diverse. When conducting ultra short-term load forecasting, it is difficult to confirm the forecasting results different types loads, degree fitting actual values low. Therefore, comprehensive model based on FCM proposed. In this study, we use DTW distance Pearson describe horizontal vertical shape similarity between curves as initial clustering centers, set number cluster centers in match type load, iterate membership center matrix using until result conditions met, output results. Additionally, calculate distribution predicted samples within corresponding range each test results, designed achieved high fit electric hydrogen loads values, providing assistance ensuring smooth operation system.

Language: Английский

---

T-RippleGNN: Predicting traffic flow through ripple propagation with attentive graph neural networks

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(5), P. e0323787 - e0323787

Published: May 28, 2025

Recently, accurate traffic flow prediction has become a significant part of intelligent transportation systems, which can not only satisfy citizens’ travel need and life satisfaction, but also benefit urban management control. However, forecasting remains highly challenging because its complexity in both topology structure time transformation. Inspired by the propagation idea graph convolutional networks, we propose ripple-propagation-based attentive neural networks for (T-RippleGNN). Firstly, adopt Ripple to capture spatial model. Then, GRU-based model is used explore through timeline. Lastly, those two factors are combined attention scores assigned differentiate their influences on prediction. Furthermore, evaluate our approach with three real-world datasets. The results show that reduces errors approximately 2.24%-62,93% compared state-of-the-art baselines, effectiveness T-RippleGNN demonstrated.

Language: Английский

---

Probabilistic prediction intervals of short-term wind speed using selected features and time shift dependent machine learning models

Mathematical Biosciences & Engineering, Journal Year: 2024, Volume and Issue: 22(1), P. 23 - 51

Published: Jan. 1, 2024

<p>Forecasting wind speed plays an increasingly essential role in the energy industry. However, is uncertain with high changeability and dependency on weather conditions. Variability of directly influenced by fluctuation unpredictability speed. Traditional prediction methods provide deterministic forecasting that fails to estimate uncertainties associated predictions. Modeling those important reliable information when uncertainty level increases. Models for estimating intervals do not differentiate between daytime nighttime shifts, which can affect performance probabilistic forecasting. In this paper, we introduce a framework short-term The designed incorporates independent machine learning (ML) models point interval during respectively. First, feature selection techniques were applied maintain most relevant parameters datasets Second, support vector regressors (SVRs) used predict 10 minutes ahead. After that, incorporated non-parametric kernel density estimation (KDE) method statistically synthesize errors (PI) several confidence levels. simulation results validated effectiveness our demonstrated it generate are satisfactory all evaluation criteria. This verifies validity feasibility hypothesis separating data sets these types predictions.</p>

Language: Английский

---

Regression-based machine learning approaches for estimating discharge from water levels in microtidal rivers

Journal of Hydrology, Journal Year: 2024, Volume and Issue: 646, P. 132276 - 132276

Published: Nov. 14, 2024

Language: Английский

---