Ben Southworth
Chris Hoder

Machine Learning

Project Proposal

24 Jan. 2012

Neural Network Snow Forecasting

Background

Even today we lack the ability to produce truly accurate weather forecasts, especially with respect to precipitation. Weather will sometimes differ dramatically from the predictions of professional meteorologists. The image to the right (HPC, 2012) is provided by the National Oceanic and Atmospheric Administration (NOAA) and demonstrates the accuracy of quantitative precipitation forecasts from three different prediction models, the North American Mesoscale (NAM), the Global Forecast System (GFS), and the Hydrometeorological Prediction Center (HPC). A threat score is a function that ranges from 0 to 1 and provides a measure of how good a prediction was, with 1 being a perfect prediction.

The major difficulty in weather forecasting lies in the chaotic and non-linear nature of the atmosphere as well as our incomplete understanding of all of the relevant physical processes (Santhanam et all, 2011). The fact that forecasts over one day in advance require considering variables over 1000 miles away makes forecasting even more difficult. Forecasting weather phenomena even several days in advance is so complex that a perfect physical understanding, and thus a perfect forecast, may be infeasible.

Traditionally, weather forecasting is done via two different methods. The first and foremost is based on numerically modeling physical simulations with computers (Santhanam et all 2011). Although there has been much work and progress in this field, precipitation forecasts are still far from perfect, primarily due to the complexity previously discussed. There is also some work using previous data to base current predictions on, independent of the physical processes. This is effectively what we aim to do. Specifically, we would like to use vast sets of historical data to help a computer ‘learn’ when it will snow based on other known weather variables, thus avoiding the necessary understanding of all relevant physical processes. Past work in this area has focused on using radar snowfall estimates to predict snowfall depth (Xiao et all, 1996). Successful results were obtained in determining snowfall depth but the prediction of snowfall itself was not considered. Additional work has been done in short-term rainfall forecasting with relatively good results in the six to twenty-four hour forecast range (French et all, 1992; McCann, 1992; Kuligowski et all, 1998; Hall et all, 1999; Manzato, 2007). Forecasting short-term snowfall with this technique is something that has not been done, but could be approached with a similar methodology.

Problem

We aim to predict the probability of snowfall in Hanover, NH in the next 24 hours at any point in time, as well as a confidence metric indicating the amount of snowfall expected to occur.

Methodology

Due to the complexity and non-linearity of atmospheric systems, previous work in machine learning-based precipitation forecast has focused on using artificial neural networks (ANN) (French et all, 1992). The ANN is an attractive choice because of its ability to determine correlations without explicit knowledge of the physics involved (Veisi and Jamzad, 2009). ANNs have been shown to produce relatively accurate predictions for short-term rainfall prediction (Manzato, 2005) as well as snowfall depth (Xiao et all, 1996). Taking this previous work as motivation for using an ANN, we aim to implement several different variations, which will include but are not limited to:

· Back propagation

· Feed forward network

· Radial basis function network

· Hopfield network

Due to the large number of variables we will be considering as well as the necessary geographic spread of measurements taken, some effort may be needed to screen input variables and determine which ones are most relevant, as well as determine how large of a geographic area is needed for our forecast.

Data

We will obtain our data from the NOAA Earth System Research Library, which provides weather data from 1948 to the 21^st century with multiple daily updates for all variables relevant to the weather and climate (PSD, 2012). Our work will focus on the most relevant variables to snowfall prediction. At the Earth’s surface this includes the mean sea level pressure (MSLP), wind direction, relative humidity, and dewpoint. At 850mb, 700mb, and 500mb (heights with respect to air pressure), the relevant variables considered will be geopotential height, relative vorticity, wind direction, relative humidity, and specific humidity. The data sets are time-series, organized in a gridded fashion with respect to latitude and longitude. How large of a spatial distribution of data we will use has not been determined, but in trying to predict snow 24 hours in advance, we certainly must consider variables within a several hundred-mile radius of Hanover, and potentially variables over a 1000-mile radius.

Timeline

- 22 Jan. – 4 Feb. 2013:

o Develop a strong understanding of neural networks and different ways of using them, e.g. back propagation, Hopfield networks, etc.

o Obtain and format data, as well as determine how to deal with the enormous amounts of data we will have, i.e. terabytes.

o Begin coding and formalize how our algorithm will learn from the data.

- 5 -19 Feb. 2013

o Complete coding for our neural network.

o Begin training from data and optimization of predictive model.

- 20 Feb. – 7 Mar. 2013

o Finish training code and test on current real time data, as well as several years of previous data we will save for testing purposes. Compare results across methods.

o Tweak algorithm, variables, and data as necessary to perfect code and improve results.

o Complete write-up on our work and results.

Sources

French, Mark N., Witold F. Krajewski, and Robert R. Cuykendall. "Rainfall forecasting in space and time using a neural network." Journal of hydrology137.1 (1992): 1-31.

Ghosh, Soumadip, et al. "Weather data mining using artificial neural network." Recent Advances in Intelligent Computational Systems (RAICS), 2011 IEEE. IEEE, 2011.

Hall, Tony, Harold E. Brooks, and Charles A. Doswell III. " Precipitation forecasting using a neural network." Weather and forecasting 14.3 (1999): 338-345.

HPC Verification vs. the models Threat Score. National Oceanic and Atmospheric Administration Hydrometeorological Prediction Center, 2012. Web. 20 Jan. 2013. http://www.hpc.ncep.noaa.gov/images/hpcvrf/HPC6ts25.gif

Hsieh, William W. "Machine learning methods in the environmental sciences." Cambridge Univ. Pr., Cambridge (2009).

Kuligowski, Robert J., and Ana P. Barros. "Experiments in short-term precipitation forecasting using artificial neural networks." Monthly weather review 126.2 (1998): 470-482.

Kuligowski, Robert J., and Ana P. Barros. "Localized precipitation forecasts from a numerical weather prediction model using artificial neural networks." Weather and Forecasting 13.4 (1998): 1194-1204.

Luk, K. C., J. E. Ball, and A. Sharma. "A study of optimal model lag and spatial inputs to artificial neural network for rainfall forecasting." Journal of Hydrology 227.1 (2000): 56-65.

Manzato, Agostino. "Sounding-derived indices for neural network based short-term thunderstorm and rainfall forecasts." Atmospheric research 83.2 (2007): 349-365.

McCann, Donald W. "A neural network short-term forecast of significant thunderstorms." Weather and Forecasting;(United States) 7.3 (1992).

PSD Gridded Climate Data Sets: All. National Oceanic and Atmospheric Administration Earth System Research Laboratory, 2012. Web. 20 Jan. 2013. http://www.esrl.noaa.gov/psd/data/gridded/

Santhanam, Tiruvenkadam, and A. C. Subhajini. "An Efficient Weather Forecasting System using Radial Basis Function Neural Network." Journal of Computer Science 7.

Silverman, David, and John A. Dracup. "Artificial neural networks and long-range precipitation prediction in California." Journal of applied meteorology 39.1 (2000): 57-66.

Veisi, H. and M. Jamzad, 2009. sc. Int. J. Sign. Process., 5: 82-92. “A complexity-based approach in image compression using neural networks.” http://www.akademik.unsri.ac.id/download/journal/files/waset/v5-2-11-5.pdf