Development of an Analogue Weather Forecast Using Machine Learning and K Nearest Neighbor Algorithms

Goals / Objectives
The cooperator will evaluate the feasibility of using the K Nearest Neighbor (KNN) for analogue weather forecast in Oklahoma, to further improve the current KNN methods, to enhance the predictability of KNN methods using Machine Learning (ML) tools, and to develop parameter datasets for use in KNN for selected OK weather stations.

Project Methods
Approximately 100 years historical climate data at four typical Oklahoma stations with annual precipitation ranging from 400 mm to 1200 mm will be used to train the K Nearest Neighbor (KNN) algorithm to detect the most predictive climate patterns in each season on each site using temperature and precipitation aggregated at various time scales for various window periods within 90 days. The Machine Learning (ML) tools will be used to optimize the KNN Feature Vector to account for differences between seasons and locations by progressively reducing a forecast error indicator using self-learning features of the Artificial Intelligence (AI). Using the ML and data mining techniques, the most predictive window period going backwards and the most predictive forecasting window period going forward will be identified for each calendar month or season. The forecast skill will be evaluated against using the present climate normal as a forecast. If the tool proven successful, it will be trained for more OK stations for decision support in Ag management.