Distilling Temporal Knowledge: A GPT-BEF Based Framework for Building Energy Consumption Forecasting

Enhancing Short-Term Electrical Load Forecasting Using SARIMA, XGBoost, LSTM, and VMD-Based Signal Decomposition

Short-term electric load forecasting is essential for stable power system management, yet remains inherently uncertain due to volatile demand patterns, weather variability, and irregular consumption behaviour. The proliferation of smart metering infrastructure has made high-resolution consumption data widely available, enabling machine learning and deep learning methods to model complex non-linear temporal patterns that conventional statistical approaches cannot capture. In this context this work proposes a framework for a comparative evaluation of four representative forecasting methods: the Seasonal Autoregressive Integrated Moving Average (SARIMA) model, Extreme Gradient Boosting (XGBoost), the Long Short-Term Memory (LSTM) neural network, and a hybrid Variational Mode Decomposition–LSTM (VMD-LSTM) model. In the hybrid approach, VMD is first adopted to decompose the load time series into several intrinsic mode functions, which are then modelled individually with LSTM to improve forecasting accuracy. RMSE, MAE, R 2 , and MAPE are used to evaluate the effect of model selection on forecasting uncertainty. The results indicate that the VMD-LSTM model exhibits the most favourable performance for the considered dataset (RMSE: 2731.95, MAE: 2190.34, R²: 0.9641, MAPE: 5.15%). However, forecasting performance may vary depending on data characteristics and modelling conditions; therefore, SARIMA and XGBoost can also provide effective results in different scenarios of short-term electricity load forecasting

Simplifying AI-Based AHU Forecasting for Sustainable Building Operation: Do Seasonal and Engineered Features Improve Prediction Accuracy?

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.

The Council of Experts: A Heterogeneous Rational-Aware Framework for Robust Energy Forecasting

Accurate forecasting of building energy consumption is a cornerstone of modern smart grid management and sustainable facility operations. However, standard deep learning approaches—such as Long Short-Term Memory (LSTM) networks and Transformers—often function as black boxes, failing to explicitly model the governing physical laws and thermodynamic constraints of building systems. This limitation frequently results in poor generalization and instability when applied to diverse building portfolios. To address this challenge, we propose a novel Rational-Aware Architecture, a heterogeneous Mixtureof-Experts (MoE) framework that decomposes the forecasting task into specialized semantic agents. The architecture comprises a Thermodynamicist (Physics-ResNet) to model enthalpy and heat transfer, a Meteorologist (WeatherCNN) to capture environmental gradients, a Manager (Time2Vec) to encode temporal cyclicities, and an Engineer (Sequence Model) to handle historical load inertia. A context-aware gating mechanism dynamically weighs these experts based on specific building characteristics. Extensive experiments on a large-scale dataset of 800 buildings from the ASHRAE Great Energy Predictor III challenge demonstrate that the proposed framework significantly outperforms standard deep learning baselines. The Rational-Aware Transformer achieved a Mean Absolute Percentage Error (MAPE) of $\mathbf{1 9. 9 7 \%}$, representing a relative error reduction of approximately 40% compared to the standard Transformer baseline (33.03%). Furthermore, the Rational-LSTM variant demonstrated exceptional stability with an $R^{2}$ score of 0.9452, effectively mitigating the volatility often observed in pure data-driven approaches. These results confirm that integrating domain knowledge into deep learning architectures yields superior robustness, precision, and interpretability for energy forecasting tasks.

Hybrid machine learning approach for short-term load forecasting in smart grids

Intelligent Load Forecasting: Leveraging Machine Learning for Accurate Demand Estimation

Accurate short-term load forecasting is essential for efficient smart grid operation and demand-side energy management. Traditional statistical models often struggle to capture nonlinear and irregular residential electricity consumption patterns. This paper proposes a hybrid deep learning architecture combining Temporal Convolutional Networks (TCN) and Long Short-Term Memory (LSTM) networks for submeterlevel load forecasting. The model uses dilated causal convolutions to extract short-term temporal features, followed by LSTM units to capture long-term dependencies in energy consumption data. Data preprocessing includes normalization and slidingwindow sequence generation to convert historical load data into supervised learning samples. The model is evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) on both synthetically generated residential load profiles and the real-world UCI Household Electric Power Consumption dataset. The Hybrid TCN-LSTM model achieves an average RMSE of 0.0428 and MAE of 0.0317 on the synthetic dataset across 250 consumers and RMSE of 0.4551 kW with MAE of 0.2842 kW on the UCI dataset. Results demonstrate that the hybrid architecture effectively captures short-term fluctuations and longterm temporal dependencies, enabling reliable load forecasting for smart grid applications.

Efficient Multivariate LSTM Forecasting of HVAC Energy Consumption in IoT-Enabled Smart Buildings

HVAC systems use up about half of the total energy in smart buildings and are a key focus of optimization. The demand of HVAC energy is very difficult to forecast with high accuracy due to the nonlinear nature of HVAC operations, high temporal variability, and interdependencies among environmental and operational variables. Traditional forecasting methods like regression based models and ARIMA often do not reflect such multivariate dependencies resulting in incompetent energy management. This paper presents a multivariate Long Short-Term Memory (LSTM) model that will be developed to learn the long-term temporal dynamics of various variables related to HVAC. The model is trained and tested on a real-world benchmark dataset, which includes 11 sensor-derived features, and uses one fully connected LSTM layer with 50 hidden units trained using the Adam algorithm. Root Mean Square Error (RMSE) and the coefficient of determination (R2) are reported per variable as measures of forecast performance. The experimental findings indicate that the model is accurate, over 90% on most variables, a fact that justifies the fact that the model is effective in overcoming the weaknesses of the traditional methods and giving accurate predictions that can be incorporated into smart building energy management systems. Further research will focus on hybrid deep learning networks and TinyML networks to run on edge devices that are IoT-enabled.