1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Title:	Wavelet-ANFIS based model for rainfall forecasting
Authors:	Seyed Ahmad Akrami (P53253)
Supervisor:	Ahmed El-Shafei, Dr.
Keywords:	Rainfall Neural networks (computer science)
Issue Date:	6-May-2014
Description:	Hujan merupakan salah satu komponen hidrologi yang rumit dalam ramalan aliran dan pengurusan air. Kini, model yang berasaskan data didapati berkesan, terutamanya dalam masalah ramalan di mana ciri-ciri proses adalah stokastik dan sukar digambarkan menggunakan model matematik yang jelas. Sebab utama keadaan tersebut adalah kerana model yang berasaskan data bergantung sepenuhnya kepada data hidro-meteorologi tanpa memberi pertimbangan secara langsung terhadap perubahan fizikal. Justeru, tidak dapat dielakkan bahawa model yang berasaskan data menyebabkan ketidakpastian kepada ramalan kerana andaian yang terlalu dipermudahkan, ketidaksesuaian data latihan, input model, konfigurasi model, dan pengalaman individu pereka model. Tesis ini membentangkan satu usaha untuk meningkatkan ketepatan ramalan hidrologi dalam tiga aspek termasuk pemilihan input, pemilihan model dan penggunaan teknik prapemprosesan data dalam ramalan. Input berbeza yang menggunakan analisis korelasi linear (LCA) diuji untuk memilih input yang optimum dalam setiap senario ramalan. Dalam senario pertama, model wakil seperti "artificial neural network" (ANN) dan "adaptive neuro-fuzzy inference system" (ANFIS) dicadangkan untuk menjalankan ramalan hujan. Kajian ini mengemukakan satu struktur ANFIS yang diubahsuai (MANFIS) setelah mempertimbang semula rekaan ANFIS konvensional dengan matlamat menjadikan teknik ANFIS lebih cekap. "Modified ANFIS" (MANFIS) adalah berdasarkan anggaran bahawa bilangan peraturan dalam FIS boleh diwakili dengan bilangan input "fuzzy sets". Dalam senario kedua, dua kaedah pra-pemprosesan data termasuk "Moving Average" (MA) dan "Wavelet Transform" (WT) dikaji selanjutnya dengan model ramalan ANFIS. Selain itu, hubungan antara "moving average" yang berbeza dan tahap penguraian oleh "wavelet" telah dibincangkan. Dua kriteria, iaitu, "Root Mean Square Error" (RMSE) dan Correlation Coefficient 2 R digunakan untuk menilai model yang dicadangkan. Perbandingan input yang berbeza menunjukkan bahawa input optimum bagi meramal hujan akhirnya ditentukan sebagai (t-1), (t-2), (t-3) dan (t-4) oleh LCA berdasarkan perbandingan antara input data yang lain. Dalam senario pertama (tanpa pra-pemprosesan data), hasil perbandingan menunjukkan bahawa "Modified ANFIS" yang berasaskan model memberi ramalan hujan yang lebih tepat; kadar kesilapan yang rendah dan kerumitan komputasi (jumlah bilangan "fitting parameters" dan "convergence epochs") yang lebih rendah berbanding dengan model ANFIS dan ANN yang konvensional. Kajian ini telah mendapati bahawa model MANFIS dengan RMSE = 0.0504 dan 2 R = 0.916 di peringkat pengesahan mempunyai prestasi yang lebih baik dalam ramalan hujan daripada model ANN dan ANFIS dengan RMSE = 0.074, 2 R = 0.788 dan RMSE = 0.051, 2 R = 0.906, masing-masing . Dalam senario kedua (dengan pra-pemprosesan data), MA dan WA digunakan. Keputusan menunjukkan bahawa terdapat beberapa persamaan antara penapis "moving average" dan penapis anggaran "wavelet" sub-siri disebabkan oleh penghapusan "noise". Selain itu, keputusan yang diperolehi menunjukkan bahawa korelasi yang tinggi dengan "moving average" 2-bulan boleh dicapai melalui "Dmey wavelet transform" berbanding dengan "Haar wavelet" untuk data hujan. Kita telah mendapati bahawa gabungan wavelet-ANFIS dengan RMSE = 0.041, 2 R = 0.982 mempunyai prestasi yang lebih baik berbanding model MA-ANFIS dengan RMSE = 0.076 dan 2 R = 0.771. Bandingan antara MA dan WA menunjukkan bahawa model yang berdasarkan penguraian "wavelet" mempunyai prestasi ramalan hujan yang lebih baik dan tepat dengan kadar kesilapan yang serendah model input ANFIS.,Rainfall is one of the most complicated hydrologic processes in runoff prediction and water management. Nowadays, data-driven models have been found effective, particularly in predicting problems where characteristics of the processes are stochastic and difficult to describe, using clear mathematical models. The most important problem is that data-driven models rely solely on previous hydro-meteorological data without directly taking the underlying physical processes into account. Hence, it is unavoidable that data-driven models introduce uncertainty to the forecasting as a result of over-simplified assumption, inappropriate training data, model inputs, model configuration, and even individual experience of modelers. The major objective of this study was to improve the accuracy of hydrological forecasting in three aspects including selection of inputs, selection of models, and the utilization of data pre-processing techniques in prediction. Different inputs, using linear correlation analysis (LCA), were first examined to select optimal inputs in each prediction mode. This study was carried out within two modes. In the first mode (without data pre-processing), representative models, such as artificial neural network (ANN) and adaptive neruo-fuzzy inference system (ANFIS) were proposed to accomplish rainfall forecasting. This study after reconsidering conventional ANFIS architecture brought up a modified ANFlS (MANFlS) structure with attention to making ANFIS technique more efficient. The Modified ANFIS (MANFIS) was based on the approximation that the number of rules in Fuzzy Inference System (FIS) could be represented by the number of input fuzzy sets. In second mode (with data pre-processing), two data pre-processing methods including Moving average (MA) and Wavelet Transform (WT) were further investigated with ANFIS forecasting model. Moreover, the relationship between different moving averages and decomposed levels by wavelets was discussed. Two criteria, namely, Root Mean Square Error (RMSE) and Correlation Coefficient were used to evaluate the proposed models. The comparison of different inputs indicated that the optimal input to rainfall forecasting was finally determined as (t-1), (t-2), (t-3) and (t-4) by LCA on the basis of comparison among the other data inputs. In the first mode (without data pre-processing), the comparison result showed that the model based Modified ANFIS achieved higher rainfall forecasting accuracy; low errors and lower computational complexity (total number of fitting parameters and convergence epochs) were compared with the conventional ANFIS and ANN models. It was found that MANFIS model with RMSE=0.0504 and R2= 0.916 in the validation stage were more accurate than ANN and ANFIS models with RMSE= 0.074, R2=0.788 and RMSE= 0.051, R2=0.906, respectively. In the second mode, (with data pre-processing), MA and WA were applied in the scheme. The results showed that there were some similarities between moving average filters and wavelet approximation sub-series filters due to noise elimination. Moreover, the results demonstrated that the high correlation with moving averages 2-months could be achieved via Dmey wavelet transform compared to Haar wavelet for rainfall data. It was found that combination of Wavelet-ANFIS with RMSE=0.041, R2=0.982 had better performance than the MA-ANFIS model with RMSE=0.076 and R2= 0.771. A comparison between MA and WA indicated that the model based on wavelet decomposition performed higher rainfall forecasting accuracy and low error as ANFIS input model.,PhD
Pages:	128
Call Number:	QA76.87.A397 2014 3 tesis
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.