Download presentation

Presentation is loading. Please wait.

Published byKelsi Broad Modified over 2 years ago

1
Virtual Weatherman: A pattern recognition approach to weather prediction Joo Hyun (Paul) Song

2
55:145 PR Final Project2 Why predict weather? Our daily activities often depend on weather Weather conditions affect transportation safety Using only current weather conditions to make plans is undesirable

3
55:145 PR Final Project3 Some history… Babylon – 650 BC China – 300 BC –Weather lores 1955 –Dawn of numerical weather prediction –Development of computers

4
55:145 PR Final Project4 Weather lores Solar halo or lunar corona is precursor to rain –60 – 70% accuracy –Movement of moisture to increasingly lower levels

5
55:145 PR Final Project5 More weather lores Red sky at night probably means good weather tomorrow –Low moisture level in air near the horizon

6
55:145 PR Final Project6 Modern weather forecasting Persistence forecasting –Simplest method of forecasting weather –Todays weather carries on to tomorrows weather (works well in steady-state weather conditions) Medium range forecasting –Analog technique Pattern recognition –Ensemble forecasting Uses lots of forecasts produced to reflect the uncertainty in the initial state of the atmosphere

7
55:145 PR Final Project7 Neural Networks Simply: variable interconnections of simple elements Formally: nonlinear function from a set of inputs to a set of outputs controlled by a vector of adjustable parameters Nonlinear Neural networks learn from examples and capture subtle functional relationships among the data even if the underlying relationships are unknown or hard to describe

8
55:145 PR Final Project8 Neural Networks (framework) Weighted combinations of activation functions –Typically chosen to be nonlinear sigmoidal functions such as logsig or tansig Set of weights that produce the best fit is estimated using gradient descent

9
55:145 PR Final Project9 Dataset 3 locations –Kuala Lumpur, Malaysia Tropical Small weather fluctuations Daily data: 10/11/2001 – 11/30/2007 –Seoul, South Korea Temperate Mild weather fluctuations with 4 distinct seasons Daily data: 1/1/1996 – 11/30/2007 (minus year 2000) –Iowa City, IA Hell on earth Meteorologists nightmare Daily data: 4/17/2002 – 11/30/2007

10
55:145 PR Final Project10 Dataset (description) Weather Underground Daily weather summary of 22 parameters –Date –Max/min/mean temperature –Wind speed –Cloud cover –Precipitation –Events –etc Hourly data also available

11
55:145 PR Final Project11 Setup MATLAB + Neural Network Toolbox Input –10 features: month, mean temp, mean dew point, mean humidity, mean pressure, precipitation, rain, thunderstorm, snow and fog –Past 3 days data –Previous years data for training –This years data for testing Neural Network –4 layer network: –purelin basis –Resilient backpropagation training function (trainrp) –1000 iterations Output –4 features: mean temp, mean dew point, mean humidity and mean pressure

12
55:145 PR Final Project12 Results (Kuala Lumpur) Actual Predicted RMSE = Mean Temp Mean Dew Point Mean Humidity Mean Atm. Pressure Mean Temp Mean Dew Point Mean Humidity Mean Atm. Pressure

13
55:145 PR Final Project13

14
55:145 PR Final Project14 Results (Seoul) Actual Predicted RMSE = Mean Temp Mean Dew Point Mean Humidity Mean Atm. Pressure Mean Temp Mean Dew Point Mean Humidity Mean Atm. Pressure

15
55:145 PR Final Project15 Results (Hell) Actual Predicted RMSE = Results (Iowa City) Mean Temp Mean Dew Point Mean Humidity Mean Atm. Pressure Mean Temp Mean Dew Point Mean Humidity Mean Atm. Pressure

16
55:145 PR Final Project16 Conclusion Neural network weather predictor performs fairly well considering small number of input features. There was slight improvement in prediction results if data for the corresponding season was used to train the system. Performance may improve with more intelligent combination of inputs (i.e. weather conditions of surrounding regions, etc). Comparison to other pattern recognition schemes such as Fuzzy set predictor may be worth investigating. Prediction of weather events using logsig/tansig activation functions would be something worthwhile to implement.

17
55:145 PR Final Project17 SORRY. NO QUESTIONS, PLEASE.

Similar presentations

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google