multi-model data fusion for hydrological forecasting linda see 1 and bob abrahart 2 1 centre for...

23
Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School of Earth and Environmental Sciences, University of Greenwich, UK

Upload: jane-johnston

Post on 13-Jan-2016

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Multi-Model Data Fusion for Hydrological Forecasting

Linda See1 and Bob Abrahart2

1Centre for Computational Geography, University of Leeds, UK

2School of Earth and Environmental Sciences, University of Greenwich, UK

Page 2: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

What is Data Fusion?• process of combining information from multiple sensors and/or data

sourcesRESULT = a more accurate solution

OR one which could not otherwise be obtained

• analogous to the way humans and animals use multiple senses + experience + reasoning to improve their chances of survival

Page 3: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Is Now a Practical Technology due to:

• provision of data from new types of sensors• development of advanced algorithms:

– Bayesian inference– Dempster-Shafer theory– neural networks

– rule-based reasoning systems • high performance computing• advances in communication

Page 4: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Areas of Use

Military applications:– automated target recognition (e.g. smart weapons)– guidance for autonomous vehicles– remote sensing– battlefield surveillance

Nonmilitary applications:– robotic navigation– law enforcement– medical diagnosis

Page 5: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Data Fusion• two main categories of data fusion:

– low level: fusion of raw information to provide an output

– higher level: fusion of raw + processed information to provide outputs including higher level decisions

• RESULT = a lack of standard terminology• differentiation by application domain, objective,

types of data/sensors used, degree of fusion

Page 6: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Data Fusion Framework

• flexible characterisation provided by Dasarathy (1997)

• divides inputs/outputs into data, features and higher level decisions– e.g. feature might be the shape of an

object + range to give volumetric size of the object

Page 7: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Simple Data-In Data-Out (DIDO) Strategy

Data Inputs

Amalgamation Technologiese.g., Bayesian inference,

neural networks,rule-based systems, etc.

Data Outputs

Page 8: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Relevance to Hydrological Forecasting

• many different hydrological modelling strategies

• may benefit from being combined

DifferentModel

Forecasts

Simple Statistics, Neural Networks

ImprovedModel

Forecast??

Page 9: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Study Areas

Two contrasting sites: Upper River Wye at Cefn Brywn (Wales,

UK)– small, flashy catchment

the River Ouse at Skelton (Yorkshire, UK)– stable regime at the bottom of a large

catchment

Page 10: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Individual Forecasting ModelsUpper River Wye River Ouse

TOPMODEL Hybrid Neural Network (HNN)Feedforward Neural Network

(NN1)ARMA[1,2] model

NN1 + weight-based pruning(NN2)

Rule-based f uzzy logic model(FLM)

NN1 + node-based pruning(NN3)

Naïve predictions

ARMA[1,2] model -Naïve predictions -

Page 11: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Simple Statistics for Combining Forecasts

#1: Arithmetic Mean– on the basis that different models might have

different residual patterns– averaging out might cancel out highly

contrasting patterns

#2: Median– might work better if the range of predicted

values are skewed

Page 12: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

NN-based Data Fusion Strategies

#MMF_1: Inputs (Skelton)

Hybrid Neural Network (HNN)Fuzzy Logic Model (FLM)ARMA modelNaïve predictions

Output (Skelton)Level at t+6

HiddenLayer

Page 13: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

NN Strategies cont’d

#MMF_2: MMF_1 but using differenced data

#MMF_3: MMF_2 + arithmetic mean of the three predictions

#MMF_4: MMF_2 + standard deviation

Page 14: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

NN Strategies cont’d

#MMF_5: Inputs from MMF_2 to predict model weightings based on best performance

e.g., if model_1 > model_2 & model_3 then

the models were assigned weights of 1, 0, 0

Page 15: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

More NN Strategies

#MMF_6: used outputs from MMF_5 + differenced predictions from the models

#MMF_7: MMF_6 + actual level at time t

Page 16: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

RMSE for Training (T) and Validation (V) DataCefn Brwyn (m3/hx104) Skelton (m)

Approach Model V(1984)

T(1985)

V(1986)

V40%

T60%

HNN - - - 0.056 0.051

FLM - - - 0.110 0.109

TOPMODEL 1.518 1.417 1.182 - -

NN1 Individual 0.611 0.461 0.582 - -

NN2 0.453 0.538 0.638 - -

NN3 0.475 0.575 0.705 - -

ARMA 0.398 0.668 0.706 0.098 0.082

PERSISTENCE 0.369 0.886 0.975 0.159 0.165

MEAN 0.424 0.528 0.516 0.086 0.087

MEDIAN 0.364 0.534 0.613 0.085 0.086

MMF_1 1.350 0.660 1.900 0.011 0.017

MMF_2 0.652 0.402 0.577 0.010 0.014

MMF_3 Multi- model 0.620 0.400 0.580 0.010 0.014

MMF_4 0.620 0.410 0.560 0.010 0.014

MMF_5 0.403 0.462 0.520 0.041 0.042

MMF_6 0.519 0.439 0.509 0.013 0.016

MMF_7 0.533 0.398 0.488 0.011 0.015

Page 17: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

MMF_2 forecasts for Skelton: 30 Oct 1991 21:00

Page 18: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

MMF_2 forecasts for Skelton: 4 Jan 1992 03:00

Page 19: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

MMF_2 forecasts for Cefn Brywn: 20 Nov 1984 06:00

Page 20: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

MMF_2 forecasts for Cefn Brywn: 27 Dec 1996 16:00

Page 21: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

MMF_2 forecasts for Cefn Brywn: 7 Oct 1994 10:00

Page 22: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

MMF_7 forecasts for Cefn Brywn: 7 Oct 1994 10:00

Page 23: Multi-Model Data Fusion for Hydrological Forecasting Linda See 1 and Bob Abrahart 2 1 Centre for Computational Geography, University of Leeds, UK 2 School

Conclusions

• can extend data fusion to many new areas including hydrological modelling

• data fusion, at the simplest DIDO level, can result in improvements in prediction but requires further testing

• also has potential relevance at higher decision making levels for flood forecasting and warning systems