Neural Network Based Seasonal Predictions of Lake-Effect Snowfall

Abstract

In the case of the city of Buffalo (New York, United States), located on the eastern shore of Lake Erie and, therefore, strongly influenced by the lake-effect, total monthly snowfall was predicted one to six months in advance. For this, neural network (NN) techniques, specifically a multi-layer perceptron, as well as a multiple linear regression (LR) model were applied. The period of analysis comprised 28 years from January 1982 to December 2009. Input data included surface air temperature; the temperature difference between the lake surface water temperature (LSWT) and the 850 hPa air temperature; the u-component of the wind (u-wind) and the v-component of the wind (v-wind), geopotential height (GPH) over Lake Erie and the surrounding regions at the 1000, 925, 850 and 700 hPa levels as well as the surface pressure; the 500 hPa GPH over James Bay, Canada; the surface pressure over the Great Plains; and the mean water temperature and LSWT of Lake Erie, as well as the amount of ice cover. Moreover, several teleconnection indices were implemented: the North Atlantic Oscillation (NAO), the Arctic Oscillation (AO), the Pacific/North American (PNA) pattern, the Southern Oscillation Index (SOI) and the Pacific Decadal Oscillation (PDO).Different lead times for the input variables were tested for their suitability. The most accurate result was obtained using the NN with an optimized one-month lead time approach (lead times varied between one and six months for the different input variables).