Integrating Fire Spread Modelling With Encoder–Decoder LSTM Forecasting for Wildfire Trajectory and Area Prediction

Abstract

Accurate short-term forecasting of wildfire trajectory, velocity, and fire-area evolution is important for early warning, simulation-based planning, and future UAV-assisted monitoring. Traditional physics-based simulators provide strong physical grounding but require extensive environmental inputs and calibration, while most data-driven methods rely on coarse satellite imagery that cannot support high-frequency frame-level forecasting. This study presents a simulation-to-learning framework that combines an Unreal Engine fire-spread model with temporal forecasting architectures. A controllable fire-growth mechanism was implemented using Niagara particle systems and Blueprint scripting to encode wind-driven advection, fuel-dependent expansion, and cone-based footprint growth, producing structured high-temporal-resolution logs of fire position, area, velocity, and environmental parameters. Using this dataset, three models were evaluated: a single-step BiLSTM-attention model for next-frame prediction, a direct encoder–decoder BiLSTM-attention model for 20-step forecasting, and a delta-based encoder–decoder model that predicts frame-to-frame state increments. All experiments used scenario-level train–validation–test partitioning before sequence generation. Results show that the single-step model provides useful immediate predictions but suffers from long-horizon drift when rolled out recursively. The direct encoder–decoder model avoids recursive feedback but struggles with absolute fire-area and trajectory reconstruction. The delta-based encoder–decoder achieved the lowest errors for the main spatial and area targets, with MAE values of 0.0145, 0.0144, and 0.0828 for fire_x, fire_y, and fire_area, respectively, in the reference test run. Multi-seed evaluation further showed that the delta formulation was the most robust model across random initialisations. Baseline and ablation studies confirmed that the delta representation was the most influential component, while BiLSTM encoding and temporal attention provided additional modelling support. The proposed framework provides a controlled foundation for fine-grained wildfire forecasting, while real UAV data integration, domain adaptation, and onboard deployment remain future work.