Open research questions in Energy Load and Power Forecasting

What the literature leaves open

• The added value of the reinforcement learning framework This work demonstrates that combining a calibrated FE digital twin, see Finite Element digital twin of the induction furnace” section, a real-time DL surrogate, see “Deep learning neural networks to exploit the digital twin” section, and a DQN agent, see “Reinforcement Learning framework to optimize energy consumption in the induction furnace” section, offers a viable route to supporting furnace energy-demand reduction while preserving billet-exit temperature-profile similarity according to the prescribed engineering criterion within the calibrated surrogate-based environment considered in this study. The FE twin, developed in Phase 2, already reproduces the resulfurized billets’ outlet profiles with RMSE < 5 °C and Pearson ρ > 0.95, providing physically grounded outlettemperature target profiles for the measured plant cases 123 used in this study. In the present implementation, the DL surrogate was trained using real plant time series from actual billets as inputs, together with the corresponding FE-simulated outlet-temperature profiles as target outputs. No Monte-Carlo-generated or otherwise synthetic operating scenarios were included in the surrogate-training dataset. However, the calibrated FE model provides a natural route for future data augmentation, since it could be driven with Monte-Carlo-generated operating conditions to create large synthetic datasets composed of realistic billet-advance and inductor-power time series paired with FE-simulated temperature profiles. Phase 3 converts the calibrated FE-based thermal representation into a millisecond-scale surrogate model by training a convolutional-recurrent network whose test-set error (≈ 9.9 °C) is close to the lower bound of the outlet-pyrometer uncertainty range. In Phase 4, this surrogate enables a reinforcement-learning loop that explores thousands of candidate heating patterns per billet, an exploration infeasible with the seven-minute FE solver. This difference in execution time is central to the practical feasibility of the framework: the FE model is not suitable for real-time use because each simulation requires around seven minutes, whereas the DL surrogate enables billet-state evaluations on a millisecond scale. In this sense, the surrogate is not only a modeling convenience but the key element that makes the optimization stage computationally tractable in practice. This surrogate-based environment satisfies the replicability requirement of industrial AI systems and balances physical fidelity and computational efficiency. Applied to the evaluated resulfurized billets, the trained DQN identified surrogate-predicted heating schedules that maintain the predicted exit-temperature profiles within, or marginally close to, the prescribed 5 °C RMSE similarity threshold and reduce surrogate-predicted energy consumption relative to GSW’s baseline. Example billets with IDs 6,211,341 and 6,211,354 illustrate how the agent fine-tunes inductor-group power pulses to correct profiles that are too cold or too hot, achieving surrogate-predicted energy reductions of 24% and 27%, respectively. A fleet-wide summary indicates that the optimized schedules reduced surrogate-predicted energy consumption for all evaluated billets, although the magnitude of these reductions should be interpreted cautiously because the schedules have not yet been re-evaluated with the FE model or validated in plant operation. From a computationalcost perspective, it is also important to distinguish between offline training and online inference. The proposed work- flow is inherently divided into offline and online stages. The FE simulations, surrogate training, and RL optimization constitute the offline phase and represent the principal computational cost of the methodology. In the present study, RL training was carried out offline and required slightly more than one hour using GPU resources.

  Optimization of the energy consumption of an induction furnace for steel billets using a reinforcement learning framework applied to a finite element digital twin · 2026 · DOI

• Several improvements can strengthen the methodology and broaden its applicability. First, the FE digital twin should be enriched by incorporating billet-absorbed electromagnetic power estimates directly into the coupled simulation and by using the full inlet-temperature time series rather than a single average value. A more physically complete twin would improve local temperature fidelity and provide richer supervisory signals for both surrogate training and reward construction. In addition, a formal sensitivity analysis should be incorporated in future work. A natural next step would be to vary key modeling parameters such as the equivalent heat-transfer coefficient, the initial billet temperature, the electromagnetic power-transfer ratio, and selected boundary conditions, and then quantify their effect on the agreement between simulations and plant measurements, for example in terms of RMSE and Pearson correlation with respect to the pyrometer data. Second, the available dataset should be enlarged using both additional production runs and synthetic operating scenarios generated with the calibrated FE model. This would reduce overfitting, improve robustness, and facilitate extension to billet families and operating regimes beyond the resulfurized grade. Future work should also include cross-validation or repeated resampling, uncertainty-aware surrogate formulations, and ablation studies to assess the robustness of the convolutional-recurrent, attention, and dimension-shuffling components under limited-data conditions. In this regard, recent reviews in physically grounded manufacturing highlight the value of simulation-generated data to mitigate data scarcity and improve the interpretability of learning-based models when physical knowledge is preserved during training (Leng et al., 2025). Third, model generalization across billet families and related operating contexts should be addressed through a transfer-learning/domain-adaptation and modulardeployment strategy. The present surrogate and DQN policy were developed for the resulfurized billet family, so direct application to other families, billet dimensions, furnace configurations, or production lines cannot be assumed. Differences in thermophysical properties, inlet-temperature conditions, billet dimensions, furnace layout, inductor con- figuration, or heating response may affect both surrogate accuracy and the optimized power schedules. However, extending the framework to other billet families or compatible operating contexts would not necessarily require rebuilding the complete pipeline from scratch. Since the furnace geometry, inductor-group configuration, state variables, power limits, and discretized action space remain unchanged within the same production line, these elements can be reused.

  Optimization of the energy consumption of an induction furnace for steel billets using a reinforcement learning framework applied to a finite element digital twin · 2026 · DOI

• Looking forward, future research should focus on extending the proposed framework to multi-bus and large- scale power systems, incorporating stochastic and robust optimization under uncertainty, and developing domain- https://doi. These include the disconnect between forecasting and economic decision-making, the underexplored lifecycle impact of AI, limitations in model generalization, challenges arising from grid heterogeneity, and the lack of policy and equity considerations.

  Towards Sustainable AI-Driven Renewable Energy Systems through Integration of Forecasting, Grid Economics and Lifecycle Assessment · 2026 · DOI

• Future work will focus on integrating additional environmental and economic vari- ables, developing dynamic temporal features that capture the exact Gregorian dates of Hajj and Umrah each year (since the Islamic calendar shifts by approximately 10–11 days annually), and designing hybrid AI frameworks that combine gradient boost- ing with deep learning approaches.

  A catboost-based framework for accurate electricity load forecasting in Saudi Arabia · 2026 · DOI

• Future work should focus on improving active and reactive power balance satisfaction in large networks. The proposed approach addresses the key limitations of conventional data-driven models; (1) poor physical consistency, (2) limited temporal awareness, and (3) weak scalability under dynamic operating conditions.

  End to end prediction of optimal power flow in sustainable power system using physics-informed spatiotemporal graph neural network · 2026 · DOI

• Furthermore, future research should investigate the integration of the Dynamic Selecting Machine [75] to further refine model selection and enhance the robustness of energy forecasting frameworks. Researchers: Future studies should investigate CNN-LSTM architectures, utilize high-resolution solar irradiance measurements, and apply Digital Twin technologies [9, 14] to enhance prediction transparency.

  Machine learning approaches for resource management and forecasting in energy consumption systems · 2026 · DOI

• detailed generalization, comparison of hybrid models to highlight the limitations of existing methods is depicted in Table 1. in robust multivariate and hyperparameter tuning. The settings, Although existing methods used in short term load forecasting has significant advancement it struggles with certain limitations as discussed below, • Traditional deep learning models struggle to capture the non-stationarity and non-linearity characteristics of load data efficiently. It lacks interpretability for complex deep learning predictions and fails to incorporate attention mechanisms to selectively highlight relevant past information. • The decomposition methods used in hybrid models are mostly univariate analysis and are sensitive to complex multivariate and non-stationary features. • The traditional hyperparameter tuning methods, including for complex manual, grid, and random, are insufficient architectures and high-dimensional spaces, as they suffer from computational overhead and noise. To address these limitations, this study proposes a unified Multivariate Variational Mode Decomposition-Temporal Fusion Transformer-GOAT Optimization Algorithm (MVMD-TFT- GOA) framework with enriching feature representation through multivariate decomposition to handle irregular and complex patterns addressing non-stationary and non-linearity, enabling granular interpretation of deep learning forecasts at multiple scales with TFT to capture temporal dependencies through its internal gating structure and multi-head self-attention mechanism, and demonstrating a robust methodology for optimizing complex neural networks with GOA to find global solution for TFT through efficient exploration and exploitation balance by avoiding premature convergence thereby pushing the boundaries of accuracy and enhancing the explainability in short-term load forecasting with SHapley Additive exPlanation (SHAP). 1.2 Significant contribution of this study The main contribution of this proposed study is summarized as follows: • A novel unified framework is developed, integrating MVMD decomposition with TFT to capture complex temporal patterns effectively. • Enhancing forecasting accuracy with low prediction errors by effectively handling non-stationary load patterns with MVMD. • Incorporation of the GOAT Optimization Algorithm (GOA) improves model stability and convergence. • Achieved reliable performance with several evaluation metrics and enhanced the interpretability with SHAP analysis. 1.3 Pipeline of proposed MVMD-TFT-GOA framework The proposed model (MVMD-TFT-GOA) schematic workflow is depicted in Figure 1. - The data preprocessing framework combines missing value assessment, interquartile range (IQR)-based statistical outlier detection, min-max normalization for feature scaling, and filter-based feature selection with PCC for significant feature ranking to ensure high-quality feature information for subsequent modeling.

  Short-term load forecasting using a metaheuristic optimized temporal fusion transformer with decomposition technique · 2026 · DOI

• The proposed model establishes a methodological frame- work that combines physical interpretability with a data assimilation mechanism suitable for continuous forecast- ing. This approach addresses two key challenges in se- quential prediction: avoiding reliance on the full histor- ical dataset through efficient updating, while maintain- ing parameter interpretability and uncertainty quantifi- cation. Similar strategies have proven valuable in high- stakes contexts such as the COVID-19 (Coronavirus Dis- ease 2019) pandemic, where estimates were recalibrated weekly across numerous locations [17]. Despite these advantages, the current physical repre- sentation relies on several simplifying assumptions that could be refined to better capture the processes govern- ing wind power generation. 15 Author et al.: Bayesian adaptive framework. NMI, 6, 10–17, 2026 Table 2: Average computational time (seconds) per estimation window for each bidding zone, together with the corresponding training window specifications. For zone 2, the window lengths correspond to the pre- expansion period (March 14, 2023 to September 9, 2024) and the post-expansion period (September 9, 2024 to January 1, 2025), reflecting the model update following the commissioning of the Skudeneshavn wind park. All runs used 50,000 MCMC iterations and were executed on a laptop equipped with an Apple M1 processor and 16 GB of RAM.

  Bayesian adaptive framework for wind power forecasting · 2026 · DOI

• The paper uses SHAP and LIME explainable AI methods to interpret energy consumption but does not compare their explanations for consistency or identify cases where XAI method choice affects policy recommendations, particularly for the conflicting GDP-energy relationships observed in developing versus developed European countries.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• LIME-based local explanations for Türkiye 2022 identify oil consumption as the strongest positive contributor to primary energy consumption, but the paper does not compare LIME results across other European countries or time periods to determine whether sectoral energy consumption patterns (oil vs. coal vs. gas dominance) are generalizable through explainable AI across heterogeneous economies.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• The SHAP analysis identifies that fossil energy share contributions become neutral (rather than suppressive) at 90-100% levels in European countries, but the paper does not explain the mechanisms of this saturation effect or validate whether this pattern holds across non-European energy systems with different infrastructure and resource constraints.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• The model demonstrates that GDP effects are heterogeneous (both positive and negative) across country-year combinations in the SHAP heatmap, but the paper does not identify threshold GDP levels or economic development stages where the relationship between GDP and primary energy consumption changes from linear to nonlinear for policy design purposes.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• Türkiye's structural energy consumption transition from coal-dominant (2014) to oil-dominant (2022) is identified through SHAP waterfall comparisons, but the paper lacks mechanistic modeling to explain why renewable energy share's suppressive effect increased while fossil energy share's suppressive effect declined—requiring causal analysis beyond XAI visualization.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• The population variable shows a weak and scattered relationship with primary energy consumption in the explainable AI model, but the paper does not decompose this relationship by economic sectors (transportation, industry, residential) where population effects might be mediated differently through energy intensity metrics.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• The SHAP analysis reveals heterogeneous and weak impacts of fossil and renewable energy shares across European countries and time periods, but the paper does not investigate which specific country characteristics (industrial composition, energy infrastructure maturity, regulatory frameworks) drive this heterogeneity in XAI feature effects.

  Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · 2026 · DOI

• 3 Future challenges and scopes Adopting AI in the smart energy industry faces a variety of bottleneck hurdles, including hazards, lack of data, poor data quality, tuning AI network parameters, inadequate technological infrastructure, shortage of skilled experts, integration difficulties, and legal and regulatory concerns.

  AI-driven approaches for optimizing power consumption: a comprehensive survey · 2024 · DOI

• To alleviate this problem, this article proposes a convolution neural network bidirectional long short-term memory (CNN-Bi-LSTM) to accurately predict the short-term three-phase load power in building the energy management system in the smart solar microgrid with the collected data from advanced metering infrastructure (AMI), which have not been investigated before.

  Three-Phase Load Prediction-Based Hybrid Convolution Neural Network Combined Bidirectional Long Short-Term Memory in Solar Power Plant · 2022 · DOI

• 56%), suggesting that static routing alone is insufficient to maintain expert diversity and effective functional specialization.

  Building-MoE: A closed-loop routing sparse mixture-of-experts time-series foundation model for building short-term load forecasting · 2026 · DOI

• However, it cannot be guaranteed that TimeGPT is always superior to benchmarks for load forecasting with scarce data, since the performance of TimeGPT may be affected by the distribution differences between the load data and the training data.

  TimeGPT in load forecasting: A large time series model perspective · 2024 · DOI

• This reinforces the importance of trend estimation filtering/decomposition methods, which are scarce and only few methods, primarily wavelet decomposition, have improved upon the forecasts generated by statistical linear models.

  Multistage optimization filter for trend‐based short‐term forecasting · 2021 · DOI

Most-cited papers in Energy Load and Power Forecasting

• Optimizing renewable energy systems through artificial intelligence: Review and future prospects · Energy & Environment · 2024 · 426 citations
• Neural basis expansion analysis with exogenous variables: Forecasting electricity prices with NBEATSx · International Journal of Forecasting · 2022 · 216 citations
• Carbon emission allowance price forecasting for China Guangdong carbon emission exchange via the neural network · Global Finance Review · 2024 · 195 citations
• Investigating the impact of data normalization methods on predicting electricity consumption in a building using different artificial neural network models · Sustainable Cities and Society · 2024 · 184 citations
• Wind and solar power forecasting based on hybrid CNN-ABiLSTM, CNN-transformer-MLP models · Renewable Energy · 2024 · 161 citations
• A hybrid deep learning model based on parallel architecture TCN-LSTM with Savitzky-Golay filter for wind power prediction · Energy Conversion and Management · 2024 · 158 citations
• A CNN-LSTM based deep learning model with high accuracy and robustness for carbon price forecasting: A case of Shenzhen's carbon market in China · Journal of Environmental Management · 2024 · 155 citations
• Hybrid deep learning models for time series forecasting of solar power · Neural Computing and Applications · 2024 · 143 citations
• Improved informer PV power short-term prediction model based on weather typing and AHA-VMD-MPE · Energy · 2024 · 137 citations
• Grid search with a weighted error function: Hyper-parameter optimization for financial time series forecasting · Applied Soft Computing · 2024 · 131 citations

Most recent work

• A Comprehensive Review of AI-based Wind Power Forecasting Over Multiple Time Horizons · Process Integration and Optimization for Sustainability · 2026
• Towards smart hotels: energy forecasting with machine learning models · Logic Journal of the IGPL · 2026
• Predictive intelligence of machine learning models for global energy perspectives and transformations towards sustainability · Energy Reports · 2026
• Toward a unified data-driven turbulence model through multi-objective learning · National Science Review · 2026
• Spatiotemporal Electric Energy Efficiency Evaluation via Hybrid Graph Neural Network and Transformer Architecture · Informatica · 2026
• ST-GNN-IOT: Spatio-Temporal Graph Neural Networks with Sensor-Proxy-Driven Dynamic Edge-Weight Modulation for Large-Scale Renewable Energy Forecasting · International Journal of Creative and Open Research in Engineering and Management · 2026
• Prediction of the Gross Electricity Generation Amount with Energy Sources Using the M5P Decision Tree Algorithm · Kırklareli Üniversitesi İktisadi ve İdari Bilimler Fakültesi Dergisi · 2026
• An accurate data-driven multi-horizon short-term electric load prediction using HODMD · Evolutionary Intelligence · 2026
• Interpreting Primary Energy Consumption in Europe and Türkiye Using Explainable Artificial Intelligence · ELECTRICA · 2026
• Multivariate forecasting of energy demand using recurrent neural networks · Advanced Engineering Informatics · 2026

Related topics in Engineering