Abstract—During emergencies such as flood and drought seasons, reservoir acts as a defence mechanism to reduce the risk of flooding and maintaining water supply. During this period, decision regarding the waterrelease is very crucial. During flood season, early waterreleasedecision should be established to prepare the reservoir for incoming in-flow. While during drought season, reservoirwater level should be maintained in order to sustain the supply and other usages. Reservoir operation during these two seasons cause conflicting decision as incoming inflow is hardly predicted. Modeling the reservoirwaterreleasedecision can be one of the solutions to this problem. The modeling is based on reservoir’s operator previous experiences when dealing with such situations. These experiences provide valuable information on the decision when the reservoirwater should be released. Temporaldatamining technique has been applied to extract temporal pattern from the reservoir operational record and neuralnetwork has been applied as the modeling tool. The neuralnetwork model was developed to classify the data that in turn can be used to aid the reservoirwaterreleasedecision. In this study neuralnetwork model 8-23-2 has produced the acceptable performance during training, validation and testing.
Anselin (2000) stated that there are three essential requirements for a well designed GIS integrated information system, including providing a data format that can be convert into other GIS styles, designing reusable components in the ‘Windows’ programming environment, and having a visual interface platform. Therefore, the second goal in this research is to develop an Artificial intelligent Spatial and temporal Information Analyst (ASIA) package, including the design of a new data format for the converting with other GIS styles, using the C++ program to build an open access artificial intelligent object module, and use the ArcView GIS software for the visualization platform. This package integrates artificial intelligent technology and spatial and temporaldata in commonly used GIS software environment. The third goal of this research is proceed with a time trend forecasts of air pollutants as a requirement for the model validation in order to prove the reasonableness and usefulness of the conceptual spatial and temporal analysis framework, and of the integrated artificial intelligent system.
Abstract. Reservoir is one of the emergency environments that required fast an accurate decision to reduce flood risk during heavy rainfall and contain water during less rainfall. Typically, during heavy rainfall, the water level increase very fast, thus decision of the waterrelease is timely and crucial task. In this paper, intelligent decision support model based on neuralnetwork (NN) is pro- posed. The proposed model consists of situation assessment, forecasting and decision models. Situation assessment utilized temporaldatamining technique to extract relevant data and attribute from the reservoir operation record. The forecasting model utilize NN to perform forecasting of the reservoirwater level, while in the decision model, NN is applied to perform classification of the cur- rent and changes of reservoirwater level. The simulations have shown that the performances of NN for both forecasting and decision models are acceptably good.
The reservoir operations are carried out by prediction, especially when the factors that influence the reservoir are known (certain), such as information about inflow and water requirements. In fact, the data on the inflow and actual water needs are not certainly known, especially in the decision of the volume of water needed to be released (release). The planning steps for operating the reservoir need to consider the two factors as uncertain numbers, because they are inherently high in the level of uncertainty. Suharyanto  shows the limitations of operating with a deterministic model. It implied that the operations would produce greater costs, especially in extreme conditions in reservoirs with a limited capacity.
Reservoirwaterreleasedecision is one of the critical actions in determining the quantity of water to be retained or released from the reservoir. Typically, the decision is influenced by the reservoir inflow that can be estimated based on the rainfall recorded at the reservoir’s upstream areas. Since the rainfall is recorded at several different locations, the use of temporal pattern alone may not be appropriate. Hence, in this study a spatial temporal pattern was used to retain the spatial information of the rainfall’s location. In addition, rainfall recorded at different locations may cause fuzziness in the data representation. Therefore, a hybrid computational intelligence approach, namely the Adaptive Neuro Fuzzy Inference System (ANFIS), was used to develop a reservoirwaterreleasedecision model. ANFIS integrates both the neuralnetwork and fuzzy logic principles in order to deal with the fuzziness and complexity of the spatial temporal pattern of rainfall. In this study, the Timah Tasoh reservoir and rainfall from five upstream gauging stations were used as a case study. Two ANFIS models were developed and their performances were compared based on the lowest square error achieved from the simulation conducted. Both models utilized the spatial temporal pattern of the rainfall as input. The first model considered the current reservoirwater level as an additional input, while the second model retained the existing input. The result indicated that the application of ANFIS could be used successfully for modeling reservoirwaterreleasedecision. The first model with the additional input showed better performance with the lowest square error compared to the second model.
Artificial NeuralNetwork is one of the computational algorithms that can be applied in developing a forecasting model for the change of reservoirwater level stage. Forecasting of the change of reservoirwater level stage is vital as the change of the reservoirwater level can affect the reservoir operator’s decision. The decision of waterrelease is very critical in both flood and drought seasons where the reservoir should maintain high volume of water during less rainfall and enough space for incoming heavy rainfall. The changes of reservoirwater level which provides insights on the increase or decrease water level that affects water level stage. In this study, neuralnetwork model for forecasting the change of reservoirwater level stage is studied. Six neuralnetwork models based on standard back propagation algorithm have been developed and tested. Sliding windows have been used to segment the data into various ranges. The finding shows that 2 days of delay have affected the change in stage of the reservoirwater level. The finding was achieved through 4-17-1 neuralnetwork architecture.
During the course of this study, several novel contributions have been applied including analysing the problems related to global modelling and conventional per- sonalised modelling for SSTD and their respective potential solutions; development of a prototype system based on the PMeSNNr framework called NeuCube M1 which comprises an encoding method employing Address Event Representation (AER) al- gorithm; a recurrent 3D SNN reservoir based on the Liquid-State Machine (LSM) concept and implementation of Spike Time Dependent Plasticity (STDP) as a learn- ing rules; an innovative input variables mapping techniques utilizing Factor Graph Matching (FGM) algorithm; a predictive personalised modelling method for early event prediction; various selections of evolving spiking neuralnetwork classifiers in- cluding a novel extended dynamic evolving spiking neuralnetwork method called deSNNs_wkNN for multi-NN classification and regression problems; a grid-search optimisation module and visualisation of the spiking network activities specifically on a group and personalised level. All these contributions are described and applied in Chapters 4, 5, 6 and 7. The methods have been applied to two real world case studies which are stroke occurrences prediction and aphid pest population predic- tion.
The multi-purpose reservoirwaterreleasedecision requires an expert to make a decision by assembling complex decision information that occurred in real time. The decision needs to consider adequate reservoirwater balance in order to maintain reservoir multi-purpose function and provide enough space for incoming heavy rainfall and inflow. Crucially, the waterrelease should not exceed the downstream maximum river level so that it will not cause flood. The rainfall and water level are fuzzy information, thus the decision model needs the ability to handle the fuzzy information. Moreover, the rainfalls that are recorded at different location take different time to reach into the reservoir. This situation shows that there is spatial temporal relationship hidden in between each gauging station and the reservoir. Thus, this study proposed dynamic reservoirwaterreleasedecision model that utilize both spatial and temporal information in the input pattern. Based on the patterns, the model will suggest when the reservoirwater should be released. The model adopts Adaptive Neuro-Fuzzy Inference System (ANFIS) in order to deal with the fuzzy information. The data used in this study was obtained from the Perlis Department of Irrigation and Drainage. The modified Sliding Window algorithm was used to construct the rainfall temporal pattern, while the spatial information was established by simulating the mapped rainfall and reservoirwater level pattern. The model performance was measured based on the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Findings from this study shows that ANFIS produces the lowest RMSE and MAE when compare to Autoregressive Integrated Moving Average (ARIMA) and Backpropagation NeuralNetwork (BPNN) model. The model can be used by the reservoir operator to assist their decision making and support the new reservoir operator in the absence of an experience reservoir operator.
3. THE DATA SOURCE 3.1 Data Base
For testing the effectiveness of the various methods, we considered as a case study the Cauvery river basin in South India. The Cauvery River extends over a length of about 1200 km, and the watershed extends over an area of more than 80 square km. The major reservoir of this river basin is the Mettur reservoir, in Tamil Nadu State. A hydrological database was developed after collating the data observations over a period of time. Some of these observations were carried out manually, while other observations were recorded using sophisticated sensors of stream and rain gauges. The release from the reservoir is a decision variable dependent on the current storage, the inflow, and to certain a extent the rainfall in the watershed. Monthly rainfall, inflow, storage and
Abstract – There has been little research into the use of hybrid neuraldatamining to improve robot performance or enhance their capability. This paper presents a novel neuraldatamining technique that analyses robot sensor data for imitation learning. Learning by imitation allows a robot to learn from observing either another robot or a human to gain skills, understand the behaviour of others and create solutions to problems. We demonstrate a hybrid approach of differential ratio datamining to perform analysis on spatio- temporal robot behavioural data. The technique offers classification performance gains for recognition of robot actions by highlighting points of covariance and hence interest within the data.
solving a problem through simulated annealing will prove incompatible with that of virtual machines or we can say that while working with virtualization of machines it will be quite incompatible with that of the features provided by the artificial intelligence techniques of neuralnetwork. We explored a read- write tool for synthesizing information retrieval systems and called it (NAZE), which will be proving that rasterization can be more realistic, efficient, can provide good certainity factor for probabilistic analysis, and replicate a relational nature of problem solving 20 . Work also shows that the
minimum support and minimum confidence. From the support and confidence information, it calculates a score for each class. However, the approach also suffers from two major drawbacks: (1) it produces a very large number of association rules, which results in a high processing overhead; and (2) its confidence-based rule evaluation measure may lead to over fitting which is a common problem in DM. When over fitting occurs, then the model may fit perfectly for training data but it is flawed with unseen data. In comparison with the association rule based predictive model, the traditional rule-based model, such as C4.5, the FOIL and RIPPER models, are substantially faster but their accuracy, in most cases, may not be as high as that of the association rule based predictive model. Considering these weaknesses Xiaoxin has proposed CPAR (classification based on predictive association), which has neither compromises the accuracy nor the speed. [Yin and Han 2003]. CPAR follows the basic idea of FOIL [Quinlan and Cameron-Jones 1993] in rule extraction. In comparison to SBA, CPAR has the following advantages: (1) It generates a much smaller set of high- quality rules with better prediction accuracy directly from the dataset; (2) It generates each rule by considering the set of already-generated rules to avoid generating redundant rules; and (3) when predicting the class label of an instance, it uses the best k rules that this instance satisfies. It uses dynamic programming to avoid making repeated calculations in rule generation. Thus it generates a smaller set of rules with higher quality and lower redundancy in comparison to the SBA method. As a result, it is much more time-efficient in both rule generation and prediction as well as achieving a high accuracy similar to the SBA method.
the problem domain is available in the form of a set of rules. The salient characteristics of their method are that it incorporates a probabilistic interpretation of the NN architecture which allows the Gaussian basis functions to act as classifiers. However, this method fails to generate appropriate rules on one of the benchmarking data namely bicycle control problems [Andrews et al 1995]. Towel and Shavilk (1993) developed a subset algorithm for rule mining from artificial NNs and this has been extended by Fu in 1994. Their method is suitable only for a simple NN structure, which has a small number of input neurons because the solution time increases exponentially with the number of input neurons. Sestito and Dillon [Sestito and Dillon 1994] demonstrated automated knowledge acquisition using multi-layered NNs in BRAINNE (Building Representations for Artificial Intelligence usingNeural Networks). The rule mining method developed by Sestito and Dillon has been tested with a number of benchmarking data sets. The basis of BRAINNE is more heuristic rather than mathematical. Lu in 1996 studied classification rule mining and reported the results in his research publication. He demonstrated the technique of mining classification rules from trained NNs with the help of neuro-links pruning. The Lu technique of rule mining needs expert knowledge to form the mathematical equations from clusters as well as solving them. Moreover, it does not guarantee a simple network structure at the end of the pruning phase. This can lead to a large set of mathematical equations for rule mining. This large set of equations can be difficult to solve simultaneously for rule mining. In a recent paper, Duch et. al. proposed a complete methodology of extraction, optimization and application of logical rules using NNs [Duch et al 2001]. In this method, NNs are used for rule extraction and a local or global minimization procedure is used for rule set optimization. Gaussian uncertainty measurements are used in this method when rules are applied for prediction.
Since all the data of the training and validation phase have to be noise free, all needed data were simulated by the Pansystem software to avoid any noise contribution. The range of parameters used to simulate the data is given in Table 1. Since the dual porosity models have more complexity and variable parameters, the training sets of these models are more than those of the homogenous models. Homogenous model ANNs were trained by 300 sets, the same as the dual porosity models. All the ANNs were validated by 40 sets for each model. To have a better evaluation, the validation sets were prepared with dierent parameters in each
Datamining is defined as the computational process of discovering patterns in large data sets involving methods at the intersection of artificial intelligence, machine learning, statistics, and systems. overall goal of the datamining process is to extract information from a data set and transform it into an understandable structure for further use. In addition to the raw analysis step, it involves database and data management aspects, data processing, model and inference considerations, interestingness metrics, complexity considerations, post-processing of discovered structures, visualization, and online updating. Datamining uses information from past data to analyze the outcome of a particular problem or situation that may arise. Datamining works to analyze data stored in data warehouses that are used to store that data that is being analyzed. That particular data may come from all parts of business, from the production to the management. Datamining interprets its data into real time analysis that can be used to increase sales, promote new product, or delete product that is not value-added to the company. Companies have been
Decision tree, support vector machine and artificial neuralnetwork prediction model for short-term rainfall forecast. The accuracy of classification of these models was compared. . Behzada et al. compare approach of ANN and SNV to predict one day lead flow runoff. By comparing the forecasting result with Support vector machine concluded that the prediction accuracy of ANN is at least as good as that of other models and in some cases better . Mishra et al (2013) presented the analysis based on datamining technique in hydrological daily discharge time series of the panchratna station in the river Brahmaputra under Brahmaputra and Barak Basin Organization in India. K-means, Dynamic Time Wrapping (DTW), and agglomerative hierarchical clustering are used to cluster and discover the discharge pattern in terms of the modelling. [8, 9, 10]. A performance comparison have been done between Support vector regression and multilayer feed-forward neuralnetwork models with respect to their forecasting capabilities. The two models have been designed to estimate the relationship between rainfall and runoff, which describes the most complex phenomenon of hydrological science . A performance comparison of three artificial neuralnetwork model has been done. These network models are: the multilayer perceptron neuralnetwork (MLPNN), the radial basis function neuralnetwork (RBFNN) and the simple neuralnetwork (SNN).The result of the study showed that performances of all three combination methods are better than that of the best individual rainfall-runoff model . Artificial neuralnetwork was one of most important widely used tool for data processing and hydrological forecasting. In this research feed forward neuralnetwork trained with Levenberg-Marquardt back propagation algorithm for rainfall forecasting. The correlation coefficient (R), Root mean square error (RMSE), Mean Absolute error (MAE) is implemented to evaluate the performances . Wang, Z. L. and Sheng, H. H. bought in the application of generalized regression neuralnetwork (GRNN) model to forecast annual rainfall. The performance of method is compared with the regression analysis method and back propagation neuralnetwork. .
The neuralnetwork is trained with Heart Diseases database by using feed forward neuralnetwork model and back- propagation learning algorithm with momentum and variable learning rate. The input layer of the network consists of 16 neurons to represent each attribute as the database consists of 16 attributes. Several neural networks are constructed with single hidden layers network and trained with heart disease dataset. A selection of maximum number of epochs is provided prior to training within which the training is expected to converge. The convergence is said to be achieved when the error between the output generated by the trained network and the actual output from the database matches within a certain error limit preset before the training. If a convergence is not achieved then training with new network configuration (i.e. hidden neuron count) is carried out. Below figure 2(a) and 2(b) shows training graph and the error graph which depicts the actual output, predicted output by the trained neuralnetwork and the absolute error difference between actual and predicted output.
DataMining is the withdrawal of knowledge from tremendously large datasets, detection of the non apparent facts that can perk up data analysis, interpretation and prediction process. Datamining is an automated discovery process of nontrivial, previously unknown and potentially useful patterns embedded in databases (e.g. ). It is the technology that enables data exploration and data visualisation of very huge databases at a high level of abstraction which requires no specific hypothesis in mind. This extraction of hidden predictive information from large datasets has great potential to help companies to focus on the most significant information in their data warehouses. Datamining is a process of following 7 D’s iterative sequence of steps (e.g. Fig. 1):
fMRI can be utilized to visualize the hemodynamic reaction in connection to neuronal activities inside the particular brain section . This hemodynamic reaction is demonstrated by the expanding measure of bloodstream to that specific certain neurons region. The oxyhemoglobin and deoxyhemoglobin attention are the changes in the elements of hemodynamic response in each brain tissue unit and in the rate of cerebral blood flow. Several fMRI techniques can be used to capture the functional signals produced by this hemodynamic response. Blood-oxygen-level-dependent (BOLD) is the most commonly used technique and is done by measuring the concentration of the oxyhemoglobin- deoxyhemoglobin element of the blood flow. The structural brain mapping, on the other hand, is provided by MRI and fMRI imaging technique shared with BOLD technique to produce brain images of a higher quality of excellent temporal as well as spatial information . Besides that, fMRI produces a functional mapping of the brain by measuring the iron element within the oxygenated blood; while structural mapping is constructed using the blood vessels dilation physiological principle appearing into the activated regions of the brain. The brain mapping produced is used to evaluate neural activity changes caused by internal or external stimuli triggers . Specifically, fMRI measures the oxygenated hemoglobin and deoxygenated hemoglobin ratios in the blood at many points of an individual with a control baseline. Neural activity is indicated by BOLD response measurements since it has been