TLDR: Researchers developed an explainable, attention-enhanced Bidirectional Long Short-Term Memory (BiLSTM) neural network for jointly forecasting temperature, solar irradiance, and relative humidity 48 hours ahead. The model, trained on 2019-2022 data and tested on 2023 data, achieved high accuracy (e.g., 1.32 °C MAE for temperature), outperforming existing benchmarks. It uses an attention mechanism to weigh important timesteps and Integrated Gradients for explainability, revealing that relative humidity and temperature are key predictors, and cyclical time features are crucial for longer horizons. This framework offers a transparent and accurate tool for applications like smart HVAC control and energy trading.
Accurate short-term weather forecasting is becoming increasingly vital for a variety of applications, from managing renewable energy sources to optimizing Heating, Ventilation, and Air Conditioning (HVAC) systems in smart buildings. Traditional methods, such as Numerical Weather Prediction (NWP) models, can be computationally intensive and lack the localized precision needed for individual building control. Machine Learning (ML) approaches have emerged as alternatives, but many existing deep learning models often focus on predicting a single weather variable at a time, overlooking the crucial interdependencies between factors like temperature, solar irradiance, and relative humidity.
Addressing these challenges, a new deep learning framework has been developed by Georgios Vamvouras, Konstantinos Braimakis, and Christos Tzivanidis. Their research introduces an explainable, attention-enhanced, Bidirectional Long Short-Term Memory (BiLSTM) neural network designed for jointly forecasting temperature, solar irradiance, and relative humidity up to 48 hours in advance. This innovative approach not only captures complex temporal patterns but also accounts for the cross-feature dependencies that are critical for precise predictions. You can read the full paper here.
The Model’s Core: BiLSTM with Attention and Explainability
The heart of this framework is a stacked BiLSTM network. Unlike standard LSTMs that process data in one direction, BiLSTMs analyze sequences in both forward and backward directions, allowing the model to integrate contextual information from the entire input window. This is particularly effective for capturing cyclical patterns and long-range dependencies in meteorological data.
A key enhancement is the integration of an attention mechanism. This mechanism allows the model to dynamically weigh the importance of each historical timestep, focusing more on recent and relevant data points when making predictions. This significantly improves forecast accuracy, especially for longer horizons, by preventing less relevant historical information from diluting the learning signal.
Furthermore, the researchers incorporated Integrated Gradients (IG) to enhance model explainability. IG quantifies the contribution of each input feature to the forecast outcomes, providing transparency into how the model arrives at its predictions. This interpretability is invaluable for decision-makers in applications like smart HVAC, where understanding the “why” behind a forecast can optimize control strategies.
Data and Training
The model was trained using historical meteorological data from 2019 to 2022, with data from 2023 reserved for evaluating its generalization capabilities. To help the model understand seasonal and diurnal cycles, cyclical time features—such as sine and cosine encodings for month and solar hour—were included. This encoding ensures that, for example, December and January are recognized as temporally close, unlike their numerical representations.
Hyperparameter optimization, a crucial step in fine-tuning deep learning models, was performed using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This advanced optimization algorithm helped determine the optimal number of past timesteps, learning rate, dropout rate, and the size and number of BiLSTM layers, ensuring the model’s peak performance.
Impressive Results and Insights
The model demonstrated remarkable accuracy on the 2023 test dataset, achieving Mean Absolute Errors (MAEs) of 1.32 °C for temperature, 31.51 W/m² for solar irradiance, and 6.71% for relative humidity. These results are competitive with, and in some cases surpass, state-of-the-art numerical weather prediction and other machine learning benchmarks.
Analysis of the attention weights revealed that the model prioritizes the most recent historical input timesteps, indicating their greater influence on immediate future predictions. For longer forecast horizons, the cyclical time features play a more stable and significant role.
Integrated Gradients provided fascinating insights into feature importance. Relative humidity emerged as the most influential predictor overall, followed closely by temperature. Solar irradiance, while important, was found to be a less reliable predictor due to its rapid and often chaotic fluctuations caused by factors like cloud cover. The cyclical features were consistently shown to provide essential information about underlying periodicities, especially for future predictions where recent meteorological values become less informative.
Also Read:
- Physics-Informed Neural Networks Enhance Smart Grid Modeling and Reliability
- Dynamic AI Ensemble Improves Aerodynamic Predictions for Automotive Design
Impact on Smart Systems and Energy Management
The demonstrated accuracy and transparency of this framework highlight its significant potential for energy-efficient building control through reliable short-term meteorological forecasting. By providing accurate and explainable joint forecasts of key weather variables, the model can support real-time system optimization in smart HVAC systems and enhance decision-making and risk management in energy trading applications.
