Machine Learning Approaches to Improve North American Precipitation Forecasts

Abstract

Numerical weather prediction (NWP) is a challenging task which involves working with micro and macro-scale spatio-temporal parameters susceptible to biases and accuracy problems. In recent years, machine learning has grown in popularity with the increasing demand in accurate weather predictions. In this study, we adopt a multimodel (ensemble) forecasting approach by collecting precipitation data from multiple NWP models of Canadian, American and European weather agencies in an effort to deploy an optimal machine learning-based weather model for real-time precipitation forecasting that will outperform the baseline. We considered 8 NWP models as inputs and combined them to create ensemble predictors using 5 different machine learning techniques along with a baseline model (mean of eight input NWP models). We demonstrate that machine learning approaches can improve upon the results of the individual NWP models. The best results were obtained by the neural-network variants with 17% improvement in the mean absolute error, 3% in the root mean squared error, 47% in the median absolute error, 5% in the maximum error, 70% in the relative bias, 41% in the false alarm ratio and 8% in the critical score index over the baseline. Neural networks also complied with the practicality constraints, with minutes of training time and near-real time prediction time.

Exploring Machine Learning Approaches to Precipitation Prediction: Post Processing of Daily Accumulated North American forecasts

“…Simple NN's are seen to be the worst performers among other ML models with an MAE of 0.78 mm/day. It is interesting to note that in our previous work [1] NN and CNN's had similar MAE of 0.79 and 0.78 mm/day respectively and GBR with MAE of 0.85 mm/day.…”

“…In this section, we will discuss our previous work [1] on Machine Learning Approaches to Improve North American Precipitation Forecasts, which serves as the foundation for the research presented in this thesis. Our approach involved employing an ensemble forecasting strategy, consolidating 6 months (Dec 2021 -May 2023) of NWP precipitation forecasts from multiple meteorological agencies in Canada, the United States, and Europe.…”

“…The results of the correlation analysis over a sample set of 3 months is shown In our previous work [1], the practical implementation of a proposed approach was one of the highest priority and the improved machine learning approach needed to be resource-light with manageable memory and processing time. In particular, we aimed for solutions which can be deployed on a desktop computer with tensorflow-capable GPU, with a memory consumption of under 16 GB and a processing time of under one hour for training and near real-time for prediction.…”

“…Data with temporal range from March 2023 up to June 2023 has been earmarked for model testing and validation. It offers a more recent dataset for evaluating the model's generalization capabilities and keep the testing season similar to our previous work in [1] for a comprehensive comparison.…”

“…The complex, multi-scale nature of precipitation phenomena, the limitations of NWP models, and inherent uncertainties in atmospheric dynamics have necessitated a search for innovative approaches to improve precipitation predictions. In this thesis, we embarked on a comprehensive exploration of the fusion of 24-h period NWP precipitation forecast using an ML and DL covering most of the continental USA and Canada with dataset provided by WeatherLogics Inc. Our preliminary work reported in [1], provided a foundation for this research, demonstrating that machine learning and deep learning algorithms can indeed enhance the accuracy and reliability of NWP precipitation predictions.…”