Search Results (4,579)

Accurate prediction of the state trend of pumped storage units (PSUs) is essential for timely anomaly detection and preventive maintenance to improve the overall economic performance of power plants. Nevertheless, the complex and time-varying characteristics of PSU vibration data increase the difficulty of accurately modeling their dynamic evolution. In response to this problem, an integrated vibration trend prediction (VTP) method for PSUs is developed by combining a temporal-enhanced generative adversarial network (TEGAN) with an improved bidirectional long short-term memory network (IBiLSTM). Firstly, TEGAN expands the original dataset by synthesizing artificial samples, thereby improving the structural diversity and representativeness of vibration data. Within TEGAN, a temporal characterization (TC) module is designed to collaboratively guide the generator and the discriminator, while a data processing module is adopted to incorporate structural priors into the learning process. Secondly, variational mode decomposition (VMD) is applied to decompose the original vibration data into intrinsic modes, followed by PSR to reconstruct the components of each modality into a higher-dimensional state space representation. Subsequently, by incorporating the proposed multi-order Kolmogorov–Arnold network (M-KAN) for high-order nonlinear fitting, IBiLSTM is employed to model each reconstructed sub-sequence. Finally, the outputs of all sub-sequences are aggregated to produce the final VTP results. The comparative evaluation verifies the advantages of the developed method in terms of prediction accuracy and robustness for PSU vibration trend forecasting. Full article

Surface deformation induced by underground coal mining is characterized by strong nonlinearity and spatial heterogeneity, which complicates early warning and ecological assessment. While InSAR-driven data assimilation models and remote sensing-based ecological indices are widely used for long-term monitoring, three fundamental limitations remain unresolved: (1) severe spatial imbalance in deformation samples biases data-driven models toward mean-reverting predictions, (2) recursive multi-step forecasting accumulates errors, leading to instability in long-horizon extrapolation, and (3) in ecological monitoring, vegetation resilience further induces a multi-year observation lag, resulting in a “pseudo-stable” bias in optical indicators. To address these issues, this study proposes an unified framework integrating multi-step deformation prediction and ecological time-lag analysis. Taking the Datong Coalfield as the study area, we utilized 231 Sentinel-1A images from March 2017 to December 2024 for SBAS-InSAR deformation inversion. A spatial stratified sampling strategy is used to extract 5894 representative points. A 24-step backward and 15-step forward windows were reconstructed to systematically compare six predictive models. Simultaneously, the Remote Sensing Ecological Index (RSEI) derived from Landsat data is used for cross-lagged analysis. The results demonstrate that: (1) The maximum deformation rate reached −276.75 mm/year, with cumulative subsidence exceeding −2000 mm. (2) At 3-step short-term forecasting, all models proved robust, with LSTM performing best (RMSE = 5.78 mm). At 15-step extreme extrapolation, however, traditional recursive models diverged significantly (Kalman, RMSE = 45.70 mm), whereas N-BEATS maintained stability and effectively mitigated temporal error cascades with an RMSE of 17.98 mm. (3) The core collapse zone exhibited concurrent ecological degradation (Lag 0), while the marginal basin presented a hidden degradation period of one to two years. It provides reliable scientific support for precise tracking and proactive safety management in complex mining areas. Full article

Accurate forecasting of electricity prices is essential for efficient operation and decision-making in deregulated power markets, particularly in the Indian electricity market, characterized by high volatility and dynamic pricing. This study presents a comparative analysis of statistical, machine learning, and deep learning methods for electricity price prediction in the Indian market, with a focus on feature dimensionality. The evaluated models include autoregressive-integrated-moving average, seasonal autoregressive-integrated-moving average, seasonal autoregressive-integrated-moving average with exogenous variables, categorical boosting, random forest, long short-term memory, bidirectional long short-term memory, and a hybrid convolutional neural network-bidirectional long short-term memory model. Historical data available on the Indian Energy Exchange webpage are deployed in this study. The models are analyzed for varying input vector sizes with features that include date, type of day, day of the week, previous day, month, and year market prices. The results indicate improved predictive performance of all model while increasing the input feature dimensionality from five to seven. The results indicate that while increasing the feature size from five to seven increases the prediction accuracy, the gains become marginal beyond seven, emphasizing the importance of feature relevance over feature quantity. From a theoretical perspective, the study highlights the dominance of short-term temporal dependencies in MCP prediction and provides empirical evidence for the point of diminishing returns in feature expansion. From a practical standpoint, the results endorse the choice of computationally efficient and interpretable models for real-world deployment. Full article

Accurate air conditioning (AC) load prediction under extreme weather conditions is critical for power grid stability and energy management. While tree-based ensemble methods such as XGBoost, LightGBM, and Gradient Boosting have become the dominant paradigm in short-term load forecasting, their effectiveness for weather-sensitive AC load prediction—particularly during extreme weather events—remains insufficiently examined. This study presents a systematic comparison of four machine learning models—Support Vector Regression with RBF kernel (SVR-RBF), LightGBM, XGBoost, and Gradient Boosting—for daily AC load estimation conditioned on measured same-day meteorological and calendar features. Based on five years of processed load and weather data from a major city in Southwest China, we construct 31 features and evaluate model performance across four scenarios: normal days, weather-extreme days, high-load P85, and high-load P90. Hyperparameters are selected within the first four years by expanding-window validation, and the fifth year is held out for testing. Our results reveal that SVR-RBF achieves an overall $R^2$ of 0.9772, substantially outperforming LightGBM (0.9091), XGBoost (0.8972), and Gradient Boosting (0.9094); paired moving-block bootstrap intervals for the tree-minus-SVR MAE differences exclude zero. The advantage of SVR-RBF is most pronounced under extreme conditions: on weather-extreme days, SVR-RBF attains $R^2=0.8316$ versus $R^2=0.39$ for the best tree model. This pattern is consistent with the smooth U-shaped temperature–load relationship captured by the RBF kernel. Additional trend-sensitivity analysis shows that annual load growth and target-level extrapolation also explain a substantial part of the tree-model degradation. Furthermore, we quantify the performance limitation of the restricted feature set: while normal-day estimation achieves $R^2=0.984$, high-load P90 days reach $R^2=0.843$. These findings support SVR-RBF as a strong baseline for measured-weather conditional AC load estimation while emphasizing that temporal shift and training-domain coverage must be considered when interpreting model differences. For energy-system applications, the lower errors under high-load and weather-extreme conditions are relevant to peak-demand assessment, reserve planning, and demand-side management. Full article

Using 14,663 firm-year observations from 2911 Chinese listed industrial firms over 2010–2022, we examine the association between earnings pressure and corporate carbon productivity. Earnings pressure is identified from firms’ proximity to analysts’ earnings-forecast thresholds, and carbon productivity is measured as main business revenue generated per unit of estimated carbon emissions. Firms near the forecast thresholds have significantly lower carbon productivity: the baseline coefficients imply differences of approximately 1.34% and 0.95% under the mean- and median-forecast definitions. The coefficients remain negative and statistically significant when firm and year fixed effects are included and standard errors are clustered by firm. The relationship also varies with abnormal discretionary-expense behavior, severe environmental risk-management failures, and general innovation input, and it is weaker among firms receiving greater analyst attention and among firms located in the Yangtze River Economic Belt. These findings extend evidence on the environmental consequences of short-term capital-market pressure to carbon-related economic efficiency. Full article

Industrial tariffs and demand-response (DR) programs pay for two quantities—daily energy and daily peak demand—yet short-term forecasting research addresses mostly the first, and rarely under the constraints of small industrial sites: strongly bimodal operation and only one to two years of records. We propose the Two-Stage Adaptive Framework (TSAF), a metering-only method that detects bimodality, classifies each day as active or inactive from its partial-day consumption, and fits a regression model to active days only; the same 21 features serve both targets. On 15 min data from ten plating factories of the Ansan Plating Industrial Complex (19 months), TSAF reaches 16.2% mean MAPE on daily energy against 58.7% for a day-ahead reference, and no deep-learning model beats the 22-parameter Ridge regressor. On daily peak, evaluated on active days, a joint multi-factory Transformer reaches 8.4% MAPE against a 14.0% constant-predictor floor, and a training-free tabular foundation model (TabPFN) reaches a comparable 7.0% without cross-factory data; intraday peak timing (≈2 h mean error) marks the limit of the meter-only design. External validation on 36 stratified UCI clients delimits the framework’s scope, and a weather ablation, specified in advance of estimation, finds no significant gain (p = 0.23). After adjusting for Stage 1 misclassification and intraday dispatch feasibility, the ten factories gain about 60 million KRW (US$46,000) per year and avoid 15.4 tCO₂ under Korean market conditions. Because TSAF needs only the smart-meter feed, power suppliers and DR aggregators can deploy it without access to customer-facility internals. Full article

In smart grids, accurate load forecasting is essential for maintaining the balance between supply and demand and ensuring the safe and stable operation of power systems. To address the strong nonstationarity of short-term load sequences, the lagged effects of meteorological factors, and variable redundancy, this study proposes a load forecast framework integrating time-lag-aware redundancy-constrained feature screening and signal decomposition. First, load and meteorological data are preprocessed. Then, key lagged meteorological features are selected by combining the Pearson correlation coefficient, mutual information, and the mRMR criterion. Furthermore, VMD is adopted to decompose the load sequence, and the decomposed modes and selected features are jointly input into the LSTM model for forecast. The results show that the proposed framework achieves high forecast accuracy, stability, and robustness under different load scenarios and electricity-consumption patterns, providing a reference for short-term load forecasting and power dispatching. Full article

Accurate sea surface temperature (SST) forecasting underpins operational oceanography and climate surveillance, yet it remains constrained by the inherent multiscale spatiotemporal variability. Existing deep learning methods often struggle to identify the physical hierarchies of SST fields, typically treating them as homogeneous inputs and thus failing to decouple large-scale coherent structures from transient features effectively. To bridge this gap, this study introduces MS-SSTNet, a scale-aware framework designed for spatiotemporal SST forecasting that leverages iterative multiscale decomposition. By iteratively distilling SST fields into hierarchical spatial modes and their principal components (PCs), the architecture facilitates a rigorous scale-decoupling representation of ocean dynamics. A dual-window temporal module is then integrated to characterize the coupling between long-term persistent trends and short-term stochastic fluctuations. Evaluations using satellite-derived SST products over the South China Sea (SCS) demonstrate that MS-SSTNet achieves robust 10th day forecast skill, yielding an overall spatiotemporal average MAE of 0.3031 °C, an average RMSE of 0.4062 °C, and an average ACC of 0.7973 across the entire 1–10 day forecast horizon. Ablation studies further underscore the indispensability of multiscale decomposition and dual-window integration in enhancing forecast fidelity across diverse spatiotemporal scales. Full article

This paper studies short-horizon trajectory forecasting through a coordinate-conditioned multiscale architecture, termed AeroMixer. The framework combines segment-wise local Cartesian re-representation, multiscale decomposition, bidirectional trend mixing, scale-specific prediction heads, and direct scale aggregation for the one-step prediction of highly dynamic aerial trajectories. The mathematical part of the paper establishes a local tangent-plane approximation bound for the geodetic-to-local map, a horizontal metric-scale characterization related to latitude-dependent distortion, and smoothness as well as spectral perturbation characterizations for the weighted and kinematically regularized objective. The analysis quantifies the local approximation error and objective regularity for the proposed representation and loss. Numerical experiments on 72,000 simulated trajectory samples compare five deep learning models under trajectory-level splits, three seeds, and batch size 256. AeroMixer achieves the lowest batch-wise and pooled global position RMSE in this matched set. Regime-wise and ablation analyses further identify the conditions under which the observed differences arise. Full article

Feature engineering has become a common step in AI-based HVAC forecasting, often involving variables calculated from raw building management system (BMS) measurements, such as temperature differences, setpoint tracking deviations, airflow balance indicators, rolling statistics, and temporal or seasonal descriptors. Accurate short-term forecasting can provide a baseline of expected operation for anomaly and fault detection and can support control optimization and operator decision making. However, real-world deployment is complicated due to differences in BMS sensor availability and data quality, as well as the preprocessing and maintenance burden associated with complex feature sets. The actual contribution of these features to the performance of AI forecasting remains underexplored, particularly for short-term prediction of air handling unit (AHU) operation. This study evaluates the impact of features on short-term AHU forecasting using three deep learning (DL) architectures: Temporal Convolutional Networks (TCNs), Long Short-Term Memory (LSTM) networks, and a hybrid CNN–LSTM model. An actual operational AHU dataset from a BMS was used to predict key operational variables, including supply and extract air temperatures, supply and extract fan operating signals, and supply air temperature setpoint-tracking error. Fan signal balance was additionally evaluated as a derived indicator calculated from the two predicted fan signals. Four input configurations were evaluated: (i) full (74 inputs), containing raw BMS measurements, short-cycle temporal variables, engineered and dynamic features, and annual-calendar information; (ii) no annual calendar (68 inputs), identical to full but excluding annual-calendar variables; (iii) raw + short-cycle temporal (20 inputs); and (iv) raw-only (12 inputs). The models used a 60-min input history to forecast the following 30-min at one-minute resolution. Persistence and Ridge models were included as reference baselines. All models were trained and tested on identical data splits and forecasting horizons to ensure a fair comparison. Each DL experiment was repeated across five independent runs, and performance was evaluated using MAE, RMSE, and R². The TCN showed the strongest overall DL performance. Raw-only achieved the highest mean R² in 11 of 15 architecture–target comparisons using just 12 inputs. The best mean DL R² ranged from 0.916 for the fan signals to 0.993 for extract air temperature. Annual-calendar features improved the TCN results but provided no consistent benefit for the LSTM or CNN–LSTM. Ridge slightly outperformed the best DL configurations for temperature-related targets, reflecting the strong short-term continuity of these signals. These findings show that recent raw BMS measurements contain most of the information needed for accurate 30-min AHU forecasting, while explicit seasonal and engineered features provide limited additional value. The resulting simpler models may in the future be used as forecasting components in predictive control and fault detection systems. However, their control and energy-saving benefits must be tested separately. Full article

Accurate short-term building energy consumption forecasting is crucial for energy-system operational efficiency and demand-side management. Existing deep learning models heavily rely on historical data, making them costly for data-scarce buildings. Furthermore, their cross-building generalization ability is limited, requiring building-specific tuning that hinders large-scale deployment. To address these issues, this study proposes a novel framework based on large language models for short-term building energy consumption forecasting. It adapts large language models through parameter-efficient LoRA fine-tuning and incorporates building domain knowledge by using enhanced feature extraction modules and prompt design. Specifically, a prompt template rich in building physical semantics was designed to leverage the abundant pre-training knowledge of the large language model (LLM). This enables the model to avoid blindly fitting the data, directly aligning with building operating rules, providing a reasonable prediction basis even with limited data, and enhancing cross-building generalization. In addition, a cross-feature attention mechanism is designed to analyze the impact of dynamic meteorological features on energy consumption, thereby improving cross-climate scenario adaptability. Finally, to handle non-typical mutations in actual building operations, depthwise separable convolutional layers decompose residual components to filter out noise while preserving key features of anomalous occupancy patterns, thereby enhancing model robustness. Experiments on five real-world datasets have shown that the proposed framework outperforms state-of-the-art baselines, achieving an average improvement of 6.06% in the MAE and 4.80% in the RMSE. More importantly, it exhibits strong few-shot learning capabilities and extends to zero-shot forecasting. By reducing data dependency and enabling cross-building generalization, the framework developed in this work achieves scalable, low adaptation cost energy consumption forecasting from single buildings to clusters. Full article

Reliable traffic flow forecasting is a core component of intelligent transportation systems; however, many current approaches are still unable to simultaneously model spatial interdependencies and long-term temporal correlations, particularly in cross-sea corridors that exhibit directional heterogeneity and pronounced temporal variability. This study aims to develop an accurate and stable traffic flow prediction framework for cross-sea corridors. To achieve this, an HGTransformer model was proposed that constructed a hypergraph from traffic nodes based on spatial proximity and correlated flow variations, and used hypergraph convolution to extract spatial node representations. These representations were then fed into a Transformer equipped with multi-head self-attention and positional encoding, enabling the model to capture global temporal dependencies in the evolution of traffic flow. Using hourly traffic flow data from the Huangmaohai cross-sea corridor, the model was tested on 1 to 4 h forecasting horizons and compared with long short-term memory (LSTM), multi-layer perceptron (MLP), random forest (RF), support vector regression (SVR), and Bayesian regression (BR) models. The proposed model achieved the best overall performance, with average mean absolute percentage error (MAPE), mean absolute error (MAE), weighted mean absolute percentage error (WMAPE), and root mean square error (RMSE) of 0.178, 13.375, 0.140, and 20.538, respectively. At the 1 h horizon, these values further decreased to 0.172, 12.678, 0.132, and 19.401, while preserving peak–valley structures more accurately under both short- and longer-horizon forecasting. The main contribution of this study lies in the systematic application and validation of the Huangmaohai Corridor dataset, including a reproducible hypergraph construction strategy tailored specifically for this particular infrastructure. Full article

Rapid urbanization has increasingly posed risks of inducing land subsidence in newly developed urban districts, posing growing threats to infrastructure safety. This study focuses on a selected rectangular area within the Shannan New District of Huainan City. Based on 94 Sentinel-1A images acquired from 2023 to 2025, the SBAS-InSAR technique was employed to obtain high-density spatiotemporal surface deformation data. The discrete monitoring points were mapped onto a 100 × 100 regular grid according to their spatial coordinates, with null-value cells retained. A spatial attention mechanism was then embedded into the Convolutional Long Short-Term Memory (ConvLSTM) network to construct a Spatial Attention–ConvLSTM (SA-ConvLSTM) model for spatiotemporal prediction, which was systematically compared with LSTM, CNN-LSTM (Convolutional Neural Network combined with Long Short-Term Memory), and standard ConvLSTM. The results demonstrate that SA-ConvLSTM achieves optimal prediction performance on the temporal hold-out test set, with a root mean square error of 2.09 mm and a coefficient of determination (R²) of 0.77. For subsidence hotspot identification, the intersection over union (IoU) reaches 0.56, and the F1-score reaches 0.72—substantially improving from 0.25 for standard ConvLSTM, confirming that the spatial attention mechanism effectively enhances the model’s capability to focus on key deformation areas. Rolling predictions of the deformation field for 2026 (12 time steps, each covering one Sentinel-1A acquisition interval of approximately 12 days) yield an estimated deformation trend ranging from −16.58 to 0.28 mm over the 12-step forecast period (approximately 144 days). This integrated framework provides a methodological reference for subsidence risk identification and mitigation in the Shannan New District. Full article

Accurate wind power forecasting is essential for renewable-energy accommodation, low-carbon dispatch, and the sustainable operation of modern power systems. However, wind power series exhibit pronounced non-stationarity, strong volatility, and multi-scale evolution, making long-term trends and short-term disturbances difficult to characterize jointly. In addition, multi-step forecasting errors tend to accumulate with increasing horizons, degrading model accuracy and stability. To address these issues, this study proposes CoFFormer, a collaborative frequency-domain-enhanced network for non-stationary wind power forecasting. The model reduces input modeling complexity, strengthens collaborative representation of heterogeneous temporal information, and suppresses output-stage error accumulation. Specifically, embedded series decomposition mitigates coupling interference between trend and fluctuation components. Differentiated temporal modeling and dynamic gating then adaptively coordinate the contributions of different feature representations, while frequency-domain residual compensation enhances the recovery of periodic structures and local oscillations. Experiments on ETTh2, wind_speed, WindPower, and Location2 demonstrate strong competitiveness across forecasting horizons. CoFFormer achieves MSE/MAE values of 0.0957/0.2238 and 0.1508/0.2889 on ETTh2 for 12- and 24-step forecasting, and 0.0617/0.1490 and 0.3838/0.3948 on WindPower for 3- and 24-step forecasting, outperforming most baselines. Ablation studies confirm the effectiveness and synergy of each component, providing an effective solution for high-accuracy multi-step forecasting of complex non-stationary wind power series. Full article

Chlorophyll-a concentration (Chl-a) is a key indicator of marine water quality, phytoplankton biomass, and aquatic ecosystem status. Multi-step forecasting is challenging because Chl-a dynamics exhibit regional heterogeneity, multi-scale variability, and complex spatial dependence. Existing models often optimize domain-averaged errors, which can mask unstable node-level predictions at high-variability nodes. We propose MS-STMoE, a heterogeneity-aware multi-scale spatio-temporal mixture-of-experts framework. It uses a Haversine-distance-based K-nearest-neighbor graph, gated multi-scale temporal convolutions with seasonal encoding to model short-term and periodic variations, and node-level sparse top-k routing that assigns differentiated expert pathways to nodes with distinct dynamics based on recent-state, temporal-mean, temporal-change, and node-prior features. Using 30-day histories to forecast the next 15 days, experiments on 300 Bohai Sea and 265 South China Sea nodes show that MS-STMoE achieves the lowest average MAE and RMSE among six recent spatio-temporal baselines. Compared with the best baseline, MAE and RMSE decrease by 5.6% and 2.7%, respectively, in the Bohai Sea, and by 11.0% and 5.0%, respectively, in the South China Sea. Step-wise and node-wise analyses indicate improved medium-to-late-horizon accuracy and modest, region-dependent reductions in high-error node tails, supporting more reliable region-aware short- to medium-term water quality monitoring. Full article