Comparison of different machine learning techniques for the source term estimation at the local scale

Mathematical models are crucial for assessing hazardous material releases' environmental and population impact, especially when source parameters are uncertain. Accidental or terrorist-related releases often lack precise source data. Meteorological observations, when available, may not accurately represent release conditions, leading to significant errors in assessing hazards downwind. A standard method for the source term estimation (STE), involves backward plume modeling from sensors to estimate release locations. However, we employed a novel approach using machine learning.

Our method simulates numerous forward plumes with varied release points and meteorological conditions to create a dataset including sensor data, meteorological variables, emission rates, and concentrations. This dataset is split into training and testing sets. We train a feed-forward neural network (NN) using sensor concentrations and meteorological data as inputs and emission rates/location as outputs. The NN's predictions are then verified against actual data.

Our results show the NN-based approach performs similarly to traditional backward-plume methods but offers quicker estimates, which is crucial for accidental releases. The computational burden of NN training can be handled offline in areas prone to releases. This approach significantly reduces estimation time, providing rapid responses to emergencies.

Although the utilization of NNs for STE yielded promising outcomes, several critical issues emerged, including overfitting of the training dataset and challenges in convergence during weight optimization, particularly with a high number of sensors. These issues may result in unrealistic outcomes, such as locating the source downwind of the sensor line. To address these challenges, we investigate alternative machine learning approaches like the k-nearest neighbors (KNN) algorithm, which might exhibits potential enhancements over the aforementioned NN-related issues.