Application of Meta-Heuristic Probabilistic Genetic Algorithms and Machine Learning Surrogate Models to Recognize Optimal Patterns of Green Stormwater Infrastructure Networks and Predict their Performance in Urban Watersheds
Abstract
We coupled numerical hydrologic models (SWMM and RHESSys) with probabilistic meta-heuristic genetic algorithms (GAs) to determine optimal locations for implementation of green stormwater infrastructure (GI) practices in urban watersheds. The GA combines hypothesis T-testing with exceedance probability as a hyper-parameter, and ɛ-progress measure for search progression and stagnation. The exceedance probability determines the level of confidence by which fitness values of new populations are considered better than the previous ones in a noisy response surface, and the ɛ-progress is used to determine the acceptable level of progress from a generation to the next. The simulation-optimization is a computationally intensive process due to its iterative nature, requiring dozens of Monte Carlo simulations for evaluation of each candidate solution in the GA population. To overcome the computational challenge and to obtain knowledge on significant features for optimal locations, the GA is merged with tree ensemble regressors and artificial neural networks as surrogates for the numerical models. The surrogate models use GA-generated archives as their training datasets to predict the distribution of probabilistic objective values, in this case peak stormwater flow reduction cost and nutrient removal cost. Preliminary results show that the addition of meta-models decreased average computational time required to reach the optimality condition across multiple GA simulations by 30%. The set of hyperparameters for each of the two meta models that result in their highest prediction accuracy, as well as their optimal performances are being determined.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFMIN11E0668H
- Keywords:
-
- 1942 Machine learning;
- INFORMATICSDE: 1952 Modeling;
- INFORMATICSDE: 1986 Statistical methods: Inferential;
- INFORMATICSDE: 3238 Prediction;
- MATHEMATICAL GEOPHYSICS