Neural Networks Based Local Weather Prediction

System

an Adam

ak, Rudolf Jak

sa and J

an Ligu

Technical University Ko

sice, Letn

a 9, Ko

sice, Slovakia

Abstract. In this paper we describe how to build a fully autonomous system for

collection, prediction and presentation of single-position meteorological data -

the local weather prediction system. By employing nonlinear statistics with neu-

ral network predictor on meteorological time-series data we were able to achieve

good results for the one-day weather prediction. This novel local statistical ap-

proach to weather prediction is different compare to standard methods which are

based on the air mass movement modelling. Main objective of this paper is to de-

scribe whole system for local weather prediction including technology, software,

methods and parameters, and also experimental results.

1 Introduction

Our local weather prediction system with neural networks is based on approximation

of weather function by black-box model from weather data collected in particular local

region. This will create the weather model for single position on the map. Weather

forecast agencies produce forecasts for bigger regions or even for continents or whole

globe. The local prediction method is not practical for agency forecasts. However, for

local applications, where we are interested in local weather course, this approximation-

based weather prediction can be useful. It is able to produce prediction with short 15

minutes period. The local weather prediction can be used as an extension and reﬁnement

of agency weather forecasts. In our papers [3][4] we documented our previous work

with local weather prediction for district heating company application. In that work

we used longer historical data and wider selection of meteorological and technology

variables for training the predictor. In this work we will address the prediction from

data from standard meteorological station.

2 Weather Prediction System

Whole system consists from meteostation, control server and the public server with

internet connection. We use standard meteostation Davis Vantage Pro 2 with simple

control panel. It is capable to measure main weather parameters as air temperature,

humidity, dew point, wind chill, heat index, air pressure, wind speed, wind degree, rain

rate and solar radiation. Important is that we can control it remotely from the Linux box.

Connection between meteostation and Linux box is provided via USB cable and the

dwt.c application. Connection between Linux box as a control server and public server

Adam

cák J., Jakša R. and Liguš J..

Neural Networks Based Local Weather Prediction System.

DOI: 10.5220/0005132400610068

In Proceedings of the International Workshop on Artiﬁcial Neural Networks and Intelligent Information Processing (ANNIIP-2014), pages 61-68

ISBN: 978-989-758-041-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Internet

Public

server

Private

server

Console

WiFi AP

MeteoStation

Fig. 1. Overall physical structure of weather prediction system, meteostation and two servers

isolated with the WiFi connection.

is based on 2.4GHz WiFi IEEE 802.11g connection secured via WPA2-PSK with AES

cipher. Public server is connected to internet. Using WiFi connection between control

and public server, these are galvanically isolated. This overall scheme is on the Figure

1. Control and public servers operate several subsystems:

1. meteostation access subsystem to measure and log data from meteostation,

2. the timer for sequential task run,

3. prediction subsystem to predict weather in near future,

4. visualization subsystem to convert numerical data to graphical data,

5. web subsystem to publish visualizations to internet,

6. and ﬁnally for right cooperation of all subsystems we need data transformation

subsystem.

2.1 Meteostation Access Subsystem

For the low-level control and data access we use free Linux application dwt.c by Con

Tassios [6]. This software is available under GPL 2 license. We wrote a patch for dwt.c

to match it better with our meteostation. We distribute this patch under the same license

as the ﬁrst author.

Patch does include changed bus speciﬁcation from RS232 to USB. It also includes

new methods for measurement of a voltage of included battery, for measurement of a

solar radiation and for measurement of evapotranspiration. Biggest change is addition

of method for diagnostic messages. These messages translate 200 forecast rules for next

4 hours created by vendor of meteostation. Measured data are written into text ﬁle line

by line, where every line is single measurement. Output format of line is of the type:

VALUE (space) VALUE. All values and their types are in the Table 1. File header is

second column starting with the character #. Measured data stream is in the ﬁrst column.

2.2 Timer Subsystem Structure

Timer subsystem structure consist of standard Linux timer cron. This tool is our syn-

chronization clock for the right timing of all other subsystems. Its setup is simple and

Table 1. Log values and their types.

VALUES MEASURAND UNITS

00:16:55 TIME HH:MM:SS

1-10-2013 DATE D-M-YYYY

2.8 OUT TEMP

◦

89 OUT HUMIDITY %

1.1 DEW POINT

◦

2.8 WIND CHILL

◦

2.8 HEAT INDEX

◦

1023.8 BAROMETER hPa

0 WIND SPEED km/h

12 WIND DEGREE

◦

0 AVG WIND km/h

0.0 RAIN TODAY mm

VALUES MEASURAND UNITS

0.0 RAIN MONTH mm

863.6 RAIN YEAR mm

0.0 RAIN RATE mm/h

0.0 RAIN STORM mm

65535 START STORM unrecognized value

622 SUNRISE HMM

1615 SUNSET HMM

0.0 SOLAR RAD W/m(2)

0.0 ET TODAY mm

0.0 ET MONTH mm

77.3 ET YEAR mm

9 FORECAST RULE 8bit number

CONSOLE

INPUT

PARSER

PRE-

DICTOR

OUTPUT

PARSER

VISUALI

SATION

WEB

TIME

CRON

TIMER

SUB-

SYSTEM

RAW

DATA

CURRENT

WEATHER

PREDICTION

ARCHIVE

INTERNET

Fig. 2. Timer subsystem structure - time ﬂow.

ﬂexible. Main activity is running every ﬁfteen minutes. This main activity starts mea-

surement, logging, parsing logs, prediction, parsing predictions, generation of visual-

izations and copying all public data to webserver.

The order of tasks is very important to generate a correct prediction. The whole

scheme of timer subsystem is on the Figure 2. The tasks sequence is deﬁned by time of

running tasks and input and output data for these tasks.

1. First running task is measurement task. This task write a new line with measure-

ment at the end of ﬁle data.log. This task is deﬁned in Meteostation Access Sub-

system section.

2. Second running task is ﬁrst parser task. This task transforms data.log ﬁle to predic-

tor input ﬁle. This transformation is needed for fast run of predictor.

3. Third running task is the predictor. This subsystem use as input parsed data.log ﬁle

to generate a prediction of outdoor temperature for next 24 hours. This prediction

is written in predictor.log at the end of ﬁle.

4. Fourth running task is second parser task. This task does transform output of ﬁrst

parser task and output of predictor task to series of ﬁles for the visualization task.

5. Visualization task is running as ﬁfth task in this sequence. Its main role is gener-

ate plots via gnuplot tool from preprocessed measurement and prediction outputs.

These plots are saved in JPEG format for simple viewing on all devices.

6. After generation of plots is running Web subsystem task for publishing outputs of

measurement, prediction and visualization. This task in ﬁrst step does upload all

relevant outputs to public webserver. Second step of this task is nonstop running

Apache webserver with uploaded data. This webserver with conﬁgured webpage

provides distribution of data on user requests. Public webpage does include simple

javascript for automatic reload after ﬁfteen minutes from last download to keep

content actualized.

2.3 Prediction Subsystem

Prediction subsystem consists of empty model of feed-forward neural network with sta-

ble conﬁguration and weights ﬁle. Neural network conﬁguration is shown in Table 3.

For the correct function of predictor this network needs trained weight ﬁle. This ﬁle

represents knowledge learned during learning phase of neural network. We used stan-

dard multilayer perceptron neural network with one hidden layer and backpropagation

learning algorithm (see [1][2]).

For training we used data obtained between 7/2011 and 3/2012. These data are pub-

lic on web page of this project [7]. Input of the prediction subsystem is time window

generated from the last 288 measurements in 15 minutes intervals but predictor does

use 144 measurements in 30 minutes intervals. This setup is copied from [5], where

additional sources of data with 30 minutes sampling were used. As the input for predic-

tor we used distance from the midnight and distance from the new year and additional

six variables: Air Temperature, Dew Point, Humidity, Barometer, Wind speed, Wind

Direction.

Output of predictor is logged into archive for later statistical processing to evaluate

quality of predictions. Also the last prediction from this ﬁle is used for the visualization

subsystem. Results of the predictor are visible on webpage of the project [7] and on the

Figure 4. These results were obtained with following settings of experiment:

1. Old predictor is winner of series of experiments based on data from KSC airport

in Ko

sice. These data are from years 2000 to 2011. From these data we used 170

Weights

Neural

Network

Predictor

Measured

data

Predicted

data

Parameter Value

Number of input neurons 866

Number of hidden layers 1

Number of hidden neurons 300

Number of output neurons 48

Activation function Sigmoid

Output transform. function ((4

x)-2)

Sigmoid parameter alfa 0.5

Bias value of neuron -1

Fig. 3. Prediction subsystem structure and conﬁguration of neural network predictor. Neural net-

work weights are trained and stored externally.

000 examples splitted to training set of 140 000 examples and testing set of 30 000

examples. The learning process was stopped after stabilization of learning error.

2. New predictor is winner of series of experiments based on data from meteostation

used in this project in Ko

sice. These data were measured from July 2011 to March

2012. All data contained 29 000 examples for learning. The learning process was

stopped after stabilization of learning error.

3. Combination of predictors is the winner of series of experiments when we used

the worst predictor from the old predictors and this predictor is re-learned using

new data from our meteostation. This relearning was stopped after stabilization of

learning error.

Predictor can be extended to predict other meteorological values as Dew Point, Hu-

midity, Barometer, Wind Speed and Direction (see [5]). All we need for this upgrade

are additional trained weight ﬁles for prediction.

2.4 Data Transform Subsystem

Data transform subsystem consists of some simple programs to convert measured data

to the format needed by predictor and visualization subsystem (input parser) and pro-

grams to convert predicted data for visualization (output parser & gnuplot) and pro-

grams to generate the archive for public.

0:30 6:00 12:00 18:00 24:00

TEMPERATURE [°C]

TIME [hours]

Comparation of predictors - TUKE 25.4.2012

Measured temperature [°C]

Old Predictor [°C]

New Predictor [°C]

Combination of predictors [°C]

Name of experiment Old predictor New predictor Combination

Sum of errors in

◦

C 224.2 291.9 125.6

Average error per neuron in

◦

C 4.67 6.08 2.62

Fig. 4. Results of the predictor on example day 25.4.2012. Errors in the table are from the same

day as the ﬁgure. Sum of errors is value after summing all 24 hours errors in one prediction.

Average of errors is value of sum divided by number of output neurons (48). For this day, the best

predictor is combination predictor. Predictor error is caused by short training set obtained from

meteostation and differences in weather between meteostation place and the airport, which was

main source of training data.

APACHE

Webpage

composition

INTERNET

current

weather

data plots

static

html/css

template

weather

prediction

plot

current

weather

iframe

Fig. 5. The web subsystem and the screenshot of webpage with included visualizations from gnu-

plot. Top-left frame on webpage is iframe with numerical weather data, embedable into external

webpages.

Input parser is simple program written in C language. Its function is to transform

measured data from archive of all measurements to time window for predictor. This

time window is written to two ﬁles. First ﬁle is hist.txt ﬁle for next processing by output

parser. Second ﬁle is priklad.txt, where the data after scaling for neural network are

written. Scaling is simple - all data are multiplied with a constant to normalize input to

interval from -2 to +2. it is also used to transform the output of predictor back to real

units.

Output parser is simple program written in C language, too. Its function is transform

output of predictor to format for visualization subsystem. The input is hist.txt ﬁle from

input parser and the output of predictor.

Current weather iframe is generated by input parser, too. This iframe contains only

last measurement and is useful as current weather numerical information. This iframe

is written in html.

Archive of measured data is split into months. It’s made from full measurement

data.log ﬁle every month after the midnight automatically by the timer subsystem. This

mini archive ﬁles are controlled for consistency by administrator of system. After the

check the ﬁles are copied to public web archive using SCP Linux command.

2.5 Visualization Subsystem

Main results of whole system - visualizations - are generated by visualization subsys-

tem. This subsystem is based on GNUPLOT program to generate plots of measured and

predicted data. As input of data we use hist.txt ﬁle to draw a current weather data plots.

But also we draw predicted data. These data are combined with current weather data

plots to visualize full time window of 72 hours back and 24 hours into future.

2.6 Web Subsystem

Webpage composition is very simple. All images and current data iframe are simply

copied into static webpage template. This is done regularly every 15 minutes as deﬁned

by the timer subsystem. Remote copying is realized using simple SCP linux command.

Webpage template includes simple code for automatic refresh of client window every

15 minutes.

3 Conclusion

Our experience with neural networks based local weather prediction and our experi-

ments with the topic show promising results - neural networks sliding window predic-

tion can be used as an alternative or a complement to standard weather forecasts.

Achieved accuracies are technologically useful and beneﬁcial, but the historical data

for the training of predictor have to be several years long. We recommend 10 years or

more for the predictor training.

Our predictor structure for one day ahead hourly prediction is based on seven me-

teorological variables inputs from past three days and encoded date/time information.

We used shallow topology with single hidden layer.

We want to emphasize the availability of local weather prediction as an alternative

and extension of regional weather forecasts. In the paper we provided all details nec-

essary to build local weather prediction system with own meteostation, predictor and

public web output. It shows relative affordability and simplicity of such system.

Acknowledgement

Research supported by the ”Center of Competence of knowledge technologies for prod-

uct system innovation in industry and service”, with ITMS project number: 26220220155

for years 2012-2015.

References

1. S. Haykin, Neural Networks (A Comprehensive Foundation), Macmillan, 1994.

2. B. Krose, and P. van der Smagt, An Introduction to Neural Networks, University of Amster-

dam, 1996.

3. J. Ko

ak, R. Jak

sa, R. Sepe

si, P. Sin

ak, Daily Temperature Proﬁle Prediction for the District

Heating Application, in Nostradamus 2013: Prediction, Modeling and Analysis of Complex

Systems, Springer, ISBN:978-3-319-00541-6, pp. 363-371, 2013.

4. J. Ko

ak, R. Jak

sa, P. Sin

ak, Prediction of Temperature Daily Proﬁle by Stochastic Up-

date of Backpropagation Through Time Algorithm, in ISF 2011 Proceedings, International

Symposium on Forecasting, International Institute of Forecasters, ISSN 1997-4124, 2011.

5. J. Adam

ak, Predikcia po

casia s vyu

zit

ım neur

onov

ych siet

ı (in slovak, En.: Weather Fore-

cast Using Neural Networks), Technical University Ko

sice, 2012.

6. Con Tassios, Davis Weather Tool - dwt.c, version 1.0.1, FREE software licensed under the

GNU General Public License (GPL) version 2, Swinburne University of Technology, Mel-

bourne, 2008.

7. J. Adam

ak, Project webpage, http://neuron.tuke.sk/adamcak/0/, 2012.