Geology Reference
In-Depth Information
with the least noise level as M = 1,056 in the study area. The studies by Entropy
theory on the data, perfectly agree with the recommendations of GT on inputs
identi
cation. But Entropy theory suggested the training data length as M = 1,040
with a maximum transinformation value of 0.869 at that point. To check the
authenticity of these two approaches the traditional approaches like data splitting and
cross-correlation approaches were applied. The traditional approach suggested that
the optimum value of training data length is in the 1,000
1,100 range but failed to
suggest the exact number as GT and entropy theory. Later analysis with AIC and
BIC agreed with the previous
-
finding in the case of input selection. But the AIC/BIC
analysis shows that there are two depressions in the information criterion curve. The
first depression is close to the data point 1,000 and the other is close to the data point
1,750. The study also aimed at checking the in
uence of data time interval on real
time data based modelling. The study performed with different frequencies (15, 30,
60 and 120 min) for different lead times. The study shows that the short lead time
forecasting (say 2 h) results is not very sensitive to data frequency but it has got
in
ed 30 min
data as the better data set for longer lead-time forecasting with minimum errors and
consistency in prediction results.
In the second section of the chapter, we have explored different data based
models and wavelet based arti
uence on longer lead-time forecasting (4 h or more). The study identi
cial intelligent models (NW, W-ANFIS and
W-SVM). The study extensively analysed the capabilities of SVM in the context of
rainfall runoff modelling. The better performance was observed in the case of the
NW model in validation and training phases in terms of minimum errors and other
statistical parameters followed by models such as the W-SVM, W-ANFIS, SVM,
LLR, ANFIS, NNARX and ANN models. The study has shown that the traditional
transfer function model ARMAX had better statistical performance than that of
ANN models. Another transfer function model, ARX, had an equally good per-
formance as that of an ANN model. Even though the NW model had better
numerical prediction results, an analysis considering different attributes like sen-
sitivity, errors, complexity has shown its weakness in making a useful model for
rainfall runoff modelling. The overall utility analysis based on utility index has
identi
ed W-SVM as the better model, followed by the SVM and LLR models. The
complex models like W-ANFIS, ANFIS and NW signi
cantly failed to show their
capabilities in making a useful model with less sensitivity and complexity.
References
1. Abrahart RJ, See L (2000) Comparing neural network and autoregressive moving average
techniques for the provision of continuous river flow forecasts in two contrasting catchments.
Hydrol Process 14:2157
2172
2. ASCE (2000) Artificial neural networks in hydrology
-
I: preliminary concepts. J Hydrol Eng
-
ASCE Task Committee Appl ANNS Hydrol 5(2):115
123
-
3. ASCE (2000) Artificial neural networks in hydrology
II: hydrologic applications. J Hydrol
-
Eng ASCE 5(2):124
137
