| Summary: | 博士 === 國立成功大學 === 建築學系碩博士班 === 99 === Many factors have to be considered in the reliability analysis of planning the regional rainwater utilization tank capacity. Based on the historical daily rainfall data, the following four analyze procedures were conducted: the regional daily rainfall frequency, the amount of runoff, the water continuity, and the reliability. The suggested designed storage capacity could be obtained according to the conditions with the demand and supply reliability. By using the output data, two different types of artificial neural network models were used to build up small area rainfall–runoff supply systems for the simulation of reliability and the prediction model. This study should have significant benefit in the future application of the instantaneous prediction or the development of related intelligent instantaneous control equipment.
In terms of the success rate of prediction as a whole, the results are about 83% for BPNN and 98.6% for RBFNN. Such result is similar to the training or testing results, indicating the model itself has good stability. Inspecting the misses or failures of 19% for BPNN and 6% for RBFNN during the artificial neural network training process of this research, it shows that the percentage of overestimation by one grade is 10.18% for BPNN and 0.93% for RBFNN, while for underestimation by one grade; the percentage is 8.80% for BPNN and 0.46% for RBFNN. In other words, even if there are mistakes in the estimation of this model, the variance in the effect is just one grade and the sample points of overestimation by one grade is similar to the underestimation by one grade. This also indicates that, based on the current model, the variance of prediction result tends to be “optimistic” but the prediction result tends to be “pessimistic”, only around 9% for BPNN and 0.5% for RBFNN, further indicating the stability of the model itself.。
Compared with three models, RBFNN was more conservative and BPNN had more optimistic estimates. Besides, we observed the predicted behavior of three models with the sensitivity analysis of the parameter between runoff coefficient and water supply and shows that the inference of water supply of RBFNN model was more conservative than the BPNN4-3-1-1 model.
Despite the fact that RBFNN was more reliable than BPNN, it still made a conservative estimate for the actual monitoring data. The error rate of RBFNN was still higher than the correction of BPNN 4-3-1-1. Although the learning speed of RBFNN was faster than BPNN, it could keep the advantage at the actual prediction by adjusting the number of hidden layers and nodes. This should have significant benefit in the future application of the instantaneous prediction or the development of related intelligent instantaneous control equipment. It is believed that changing different transfer functions (e.g., binary, logistic, sigmoid, etc.) for the simulation or adopting different series of ANN structure for model construction analysis is necessary in the future.
|
|---|
