MONTHLY RESERVOIR INFLOW FORECASTING IN THAILAND: A COMPARISON OF ANN-BASED AND HISTORICAL ANALOUGE-BASED METHODS

Transcription

1 Annual Journal of Hydraulic Engineering, JSCE, Vol.6, 5, February MONTHLY RESERVOIR INFLOW FORECASTING IN THAILAND: A COMPARISON OF ANN-BASED AND HISTORICAL ANALOUGE-BASED METHODS Somchit AMNATSAN, Yoshihiko ISERI, Aki YANAGAWA, Sayaka YOSHIKAWA, Kaoru KAKINUMA, Shinjiro KANAE 4 Student Member of JSCE, M. Eng., Ph.D. Research Student, Department of Civil Engineering, Graduate School of Science and Engineering, Tokyo Institute of Technology (-- O-okayama, Meguro-ku, Tokyo, 5-855) Member of JSCE, Ph.D., Researcher, Department of Civil Engineering, Graduate School of Science and Engineering, Tokyo Institute of Technology (-- O-okayama, Meguro-ku, Tokyo, 5-855) Member of JSCE, Ph.D., JSPS Research Fellow, Department of Civil Engineering, Graduate School of Science and Engineering, Tokyo Institute of Technology (-- O-okayama, Meguro-ku, Tokyo, 5-855) 4 Member of JSCE, Ph.D., Professor, Department of Civil Engineering, Graduate School of Science and Engineering, Tokyo Institute of Technology (-- O-okayama, Meguro-ku, Tokyo, 5-855) Accurate forecasting of reservoir is essential for effective reservoir management. In this study, artificial neural network (ANN)-based models and forecasting methods based on historical analogues were used to forecast the monthly reservoir s of Sirikit Dam in the Nan River Basin of Thailand. Incorporating sea surface temperatures and ocean indices in the ANN model significantly improved the forecasting result. The wavelet decomposition of inputs before they were fed into the ANN model also improved the forecasting result. The variation analogue forecast produced the best result among the forecasting methods investigated, based on historical analogues. It was also superior to other forecasting methods when forecasting extreme values. Key Words : Reservoir forecasting, Artificial neural network, Wavelet artificial neural network,,, Variation analogue forecast. INTRODUCTION The effective management of water resources requires integration of both structural and nonstructural measures. Reservoirs are generally used in water resource management, and can serve many purposes, for example, as a water supply and/or for agriculture, navigation, and flood prevention and mitigation. However, the ability of reservoir management programs to meet their intended purposes, particularly for flood, drought, and water shortage protection and mitigation, strongly depends on reservoir information. Forecasts of reservoir s have to be accurate and timely enough for reservoir managers to operate effectively. Due to the importance of reservoir forecasting, several models and methodologies have been developed and applied for forecasting ). Although rainfall-runoff models are usually used to estimate reservoir s, they are of limited practicality in most contexts due to their complexity and data availability requirements. Moreover, the models are difficult to review or re-calculate when the required data are changed. These limitations make artificial neural network (ANN)-based and historical analogue-based models better options for the prediction of reservoir s. The use of the ANN model in several studies has indicated that it can successfully predict both short-term and longterm reservoir ). For seasonal reservoir prediction, the use of climatic indices as additional information to a historical time series in an ANN improves predictions compared to the same model using only reservoir s ). Moreover, several hybrid ANN models have been proposed to improve prediction results. The integration of wavelet analysis, i.e., the wavelet-based artificial neural network (WANN) model, is a hybrid ANN model that significantly improves the accuracy of prediction 4). The persistence forecast and the

2 weighted mean analogue forecast are forecasting methods based on historical flow analogues. They have been proven to accurately forecast -month and -month river flows at 9 stations in the United Kingdom 5). This study used the ANN, WANN, persistence forecast, and weighted mean analogue forecast to forecast the reservoir of Sirikit Dam in Thailand. In addition to these four methods, a new method of forecasting, the variation analogue forecast, was developed and used to forecast the reservoir of the dam. The methods used in this study are listed in Table. The forecasting performances of the methods used were compared using four indicators: the root mean square error (RMSE), correlation, efficiency index, and coefficient of determination.. STUDY AREA AND METHODS () Study Area and Data Sirikit Dam is located in the Nan River Basin of Thailand as shown in Fig.. It is the largest earthfilled dam in Thailand, with a catchment area of, km and a maximum storage of,64,, m. The main uses of the dam are flood prevention, the supply of water for domestic use, ecological conservation, agriculture, industrial abstractions, fishing, and as an important tourist attraction. The sources of data used in this study are listed in Table. The full details of the datasets cannot be provided here due to limitations in the Table List of forecasting methods used in this study Type of method ANN-based Historical analogues-based Method Artificial neural network (ANN) model Wavelet artificial neural network (WANN) model Variation analogue forecast length of the paper. The sea surface temperature (SST) of the Pacific Ocean, South China Sea, and Andaman Sea were areally averaged from the Extended Reconstructed Sea Surface Temperature (ERSST) version b dataset, on a -degree global grid resolution. The data used were obtained for the period from January 974 to December 4. For the ANN and WANN forecasts, data from January 974 to December 4 were used for training, from January 5 to December for validation, and from January to December 4 for testing of the models. For the persistence, weighted mean analogue, and variation analogue forecasts, the data from January 974 to December 4 were used as the historical analogues for the forecasting of from January 5 to December 4. This forecasting period corresponded to the validation and testing periods in the ANN and WANN models. () ANN The multilayer perceptron (MLP) feedforward network with one hidden layer was adopted. The network was trained in a supervised manner with an error back-propagation algorithm. Its structure is shown in Fig.. Several trials of input patterns were performed in the forecasting of reservoir s using the ANN model. The first experiment used the reservoir itself as the input of the model. Other experiments incorporated SSTs and ocean indices as Table Sources of data used in this study Data used Sea surface temperature (SST) Ocean index Monthly Reservoir SST regions/ocean index name NINO + NINO NINO.4 NINO 4 Pacific Ocean (-5 N, -45 E) South China Sea (4-8 N, -7 E) Andaman Sea (7- N, E) Southern Oscillation Index (SOI) Dipole Mode Index (DMI) - Source US National Oceanic and Atmospheric Administration (NOAA) Japan Agency for Marine- Earth Science and technology (JAMSTEC) Electricity Generating Authority of Thailand (EGAT) Input signal Output signal Input Layer Hidden Layer Output Layer Fig. Location of Sirikit Dam Fig. Structure of the neural network used in this study

3 inputs of the ANN model for the reasons explained below. The sea surface can supply water vapor to the atmosphere. Therefore, variation in atmospheric water vapor could be related to SST anomalies. As a result, the variability of rainfall and reservoir s could be associated with SST anomalies. Some previous studies have predicted reservoir using climate indices derived from SST ), and several studies have indicated that SSTs and ocean indices are associated with the seasonal and inter-annual climate of Thailand 6). To design suitable inputs for the ANN model, an autocorrelation analysis between the different lag versions of s was performed in the first experiment. In other experiments, cross correlation analyses between the different lag versions of other inputs and the s were performed. Trials were also performed in which the activation function, learning rate, number of hidden neurons, and momentum of the ANN network were changed to obtain the best forecasting result. () WANN The wavelet analysis was used as the data preprocessing technique to improve the accuracy of the ANN model. After obtaining the best forecasting result for each input dataset, the original input data were decomposed into their detail (high frequency) and approximation (low frequency) components by a discrete wavelet transform. The Haar wavelet, the simplest and oldest of all wavelets, was used. The simplicity of this wavelet function eased the decomposition process and consequently supported practical implementation. Only one level of decomposition was used in this study. After the decomposition, the decomposed data were used as the input to the ANN model. The architecture of the WANN model used in this study is shown in Fig.. (4) and weighted mean analogue forecast These two forecasting methods are based on the historical analogue of reservoir s. The forecasts were obtained by first calculating the reservoir anomalies. In the calculation, monthly reservoir s were transformed to a log form to ensure that the distribution of the s were similar to a normal distribution, and when Input Time Series Wavelets High Pass Filter Wavelets Low Pass Filter Detail Component Approximation Component Input Layer Fig. Architecture of wavelet-based artificial neural network (WANN) model... Hidden Layer Output assessing the similarity of the analogues to the recent past, the very highest s become less outstanding. After a log transformation, standardized reservoir anomalies were calculated for use in the analysis as follows. For each of the calendar months (mon), the mean reservoir (m mon) and standard deviation (s mon) were calculated from the log-transformed monthly reservoir (q t). Then a series of standardized monthly anomalies (a t) were calculated as: aa tt = qq tt mm mmmmmm ss mmmmmm, () where t denotes the serial number of the month, starting from January 974, and mon refers to the calendar month to which t pertains. Forecasts could be made once the observed data for the latest month was received, as explained below. a) The persistence forecast was obtained by comparing the observed standardized of the most recent month of the current year and the historical years. The historical year with a standardized closest to that of the current year was selected as the potential analogue. The standardized of the next month of the current year was forecasted to persist as the standardized of the selected historical year. b) For the weighted mean analogue forecast, the monthly anomalies of the most recent past months were compared to all possible historical sequences of anomalies covering the same months of the year. From this annual series of potential analogues based on the RMSE, the N ana historical analogues most similar to the recent past were selected. Then the inverse of these RMSEs were used to weight the anomalies in the months following the analogues, to form the weighted mean analogue forecast. The RMSE was calculated for each potential analogue in the observed record as follows: RRRRRRRR = DD aa DD pp (kk) aa rr (kk) aaaaaa aaaaaa KK=, () where a p(k) is the anomaly for each month k in the potential analogue of duration D ana, and a r(k) is the corresponding anomaly in the recent past. The RMSEs for the selected N ana analogues were used to calculate the weight, w, for each analogue as follows: ww(nn) = / NN aaaaaa RRRRRRRR(nn) nn=, () RRMMMMMM(nn) where n =,, N ana is the rank of the ordered RMSEs (the potential analogue a p(n) with the smallest RMSE has rank n = ). The weighted mean forecast

4 anomalies, a f(m) for each month m =,, D f in the forecast duration, D f, form the last part of the constructed analogue, a c, and were calculated as: N a f (m) = a c (D ana + m) = ana n= w(n)a p,n (D ana + m), (4) where a p, n is the vector of anomalies for the potential analogue with rank n. The D ana, N ana, and D f were set to 5, 5, and, respectively. (5) The variation analogue forecast In addition to the above methods of forecasting, a new forecasting method, the variation analogue forecast, was developed and used to forecast the reservoir. This method compares the variation in the standardized s instead of comparing the standardized as in the persistence and the weighted mean analogue forecasts. The comparison of the variation in standardized s is similar to a comparison of the displacement of objects subjected to a force. If objects have the same properties and are subjected to the same force, the displacement of those objects will be the same regardless of their initial location. Building on this concept, if the displacement of one of these objects is known, it is possible to predict the displacement of the other objects. Applying this to forecasting, if the variation in in the current month of the current year is similar to the variation in in the same month of the historical year, the s of both years must occur due to similar forcing factors. If it is assumed that this forcing factor persists, the variation in in the next month of the current year can be forecasted from the variation in of the historical year. Using this method, reservoir s were standardized as in the persistence and weighted mean analogue forecasts. Then the variation in the standardized s between successive months were calculated and used in the forecasts as described above.. RESULTS AND DISCUSSION () Forecasting results using the ANN model Table Input parameters of the artificial neural network (ANN) model that produced the best forecast Input parameter SST regions/ocean index name Lag used (month) NINO + 5, 7, 8 NINO 4, 6, 7 Sea surface NINO.4 5, 5, 6 temperature (SST) Pacific Ocean 6, 7, 8 South China Sea 6, 8, 9 Andaman Sea 7, 8, 9 Ocean index SOI 5 DMI 6 Reservoir -, Total number of input parameters After several trials, the best forecasting result was obtained from the model with input parameters as shown in Table. The activation function of this model in both the hidden and output layers was the hyperbolic function. The number of hidden neurons, learning rate, and momentum that provided the best result were,., and -.5, respectively. The performance indicators of this forecasting result are shown in Table 4. The value of all indicators for the ANN model indicated good performance for all forecasting periods. However, this forecasting was performed with the original data, and hence seasonality may have affected the results. To test this effect, another forecast was obtained using the standardized s and SST anomalies. The was standardized in the same way as in the persistence forecast. A correlation analysis between standardized s and SST anomalies was performed for the selection of suitable inputs. The best performance indicators for this forecast are shown in Table 4. The indicators confirmed that the ANN model was not good at predicting anomalies. A good forecast of the original was obtained due to the seasonality that was present in the and SST data. () Forecasting results using the WANN model Based on the forecasting results using the ANN model that successfully predicted only the original, the forecast using the WANN model was done with the original data only. After obtaining the best forecast from the ANN model, all input parameters were decomposed into their detail (high frequency) and approximate (low frequency) components. Then all decomposed components were Table 4 Performance indicators of the artificial neural network (ANN) model that produced the best forecast Model Original Standardized period RMSE r EI CD RMSE r EI CD Training Validation Testing RMSE, root mean square error; r, the correlation; EI, efficiency index; CD, coefficient of determination. The details of these indicators can be found in many literature sources, e.g., Amnatsan et al. 7). Table 5 Performance indicators of the wavelet-based artificial neural network (WANN) model Model period Model performance indicators RMSE r EI CD Training Validation Testing

5 fed into the neural network model. The performance indicators of the model with decomposed inputs are shown in Table 5. There was very good improvement in all indicators, except the coefficient of determination for the testing period. () Forecasting results using persistence and weighted mean analogue forecasts The persistence forecast and the weighted mean analogue forecast are based on historical analogues. Fig. 4 shows the plot of standardized s for forecasting in January 6. The performance indicators of these two methods in the forecasting of standardized s from January 5 to December 4 are compared in Table 6. The weighted mean analogue forecast provided the best results. (4) Forecasting results using the variation analogue forecast The variation analogue forecast is also based on historical analogues. However, this method compares the differences between successive standardized s to find the historical analogue with the most potential. Fig. 5 shows a plot of the variation value and the standardized value of s from February to January of the following year. Assuming that the most current month was December 5, we could forecast the in January 6. The variation from March to January of the historical years was plotted alongside the variation from March to December of 5. Then the most similar variation analogue was selected by comparing the variation vector from November to December. In this example, Standardized Inflow Standardized Inflow Jan 6_forecasted Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Jan 6_forecasted Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Fig.4 Plot of standardized s used for forecasting in January 6 (upper; persistence, lower; weighted mean analogue) the most similar analogue was the plot of year , as shown in Fig. 5. Thus, the variation in January 6 was calculated from the variation in December 99 and January 994. Then the forecasted standardized value for January 6 was calculated from this forecasted variation value, as shown in Fig. 5. The plot of standardized s from October 5 to January 6 was very similar to the plot from October 99 to January 994. The advantage of the variation analogue forecast is its ability to locate similar patterns between events although they happened in different zones. The performance indicators of this method in forecasting standardized s from January 5 to December 4 are shown in Table 7. A time series plot and a scatter plot between the forecasted and observed standardized s of the methods based on historical analogues are shown in Fig. 6. The variation analogue method had better forecasting performance than the other methods. The performance indicators calculated from the in its normal form also indicated that the variation analogue forecast could better predict the reservoir than the ANN and WANN models. The coefficient of determination was close to, which indicates that it could successfully predict extreme values. Standardized _Standardized Jan 6_Standardized_forecasted 5-6_Variation standardized Fig.5 Plot of the variation and the standardized values of s used for forecasting in January 6 Table 7 Performance indicators of variation analogue forecast Performance indicators 5-6_Standardized _Variation standardized Jan 6_Variation standardized_forecasted Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Table 6 Performance indicators of persistence and weighted mean analogue forecasts Calculated from standardized Calculated from normal RMSE r EI CD RMSE r EI CD Forecasting method Performance indicators RMSE r EI CD Persistence Weighted mean analogue Variation standardized

6 Standardized Standardized Standardized Forecasted Observed Jan 5 Jul 5 Jan 6 Jul 6 Jan 7 Jul 7 Jan 8 Jul 8 Jan 9 Jul 9 Jan Jul Jan Jul Jan Jul Jan Jul Jan 4 Jul 4 Month-Year Forecasted Observed Jan 5 Jul 5 Jan 6 Jul 6 Jan 7 Jul 7 Jan 8 Jul 8 Jan 9 Jul 9 Jan Jul Jan Jul Jan Jul Jan Jul Jan 4 Jul 4 Month-Year Variation analogue forecast Forecasted Observed Jan 5 Jul 5 Jan 6 Jul 6 Jan 7 Jul 7 Jan 8 Jul 8 Jan 9 Jul 9 Jan Jul Jan Jul Jan Jul Jan Jul Jan 4 Jul 4 Month-Year Fig.6 Plots of the relationship between forecasted and observed standardized s for the persistence (upper), weighted mean analogue (middle), and variation analogue (lower) methods 4. CONCLUSION Forecasted standardized Forecasted standardized Forecasted standardized Observed standardized Observed standardized Variation analogue forecast Observed standardized We used several methods to forecast the reservoir s of Sirikit Dam in Thailand. Incorporating SSTs and ocean indices in the ANN model significantly improved the forecasts. The wavelet decomposition of the inputs before they were fed into the ANN model also improved the forecasting performance. However, the good forecasting performance of the ANN model was due to the seasonality in the s and SSTs. The performance indicators of the ANN model had low values when using data anomalies to predict reservoir anomalies. The variation analogue forecast produced the best forecasts among the methods based on historical analogues. The forecasts obtained by the persistence and weighted mean analogue methods were obtained by comparing the values of the standardized s. The persistence forecast performed well only if the patterns for the historical and current years were similar in a nearby zone. The forecast of the weighted mean analogue method depends on the RMSE between the s of the historical and current years. This means that it will include all nearby patterns, even if they are not similar to the pattern of the current year. On the other hand, the variation dialogue method obtains a forecast by comparing the variation in the reservoir s, and therefore it could detect similar historical patterns even if they occurred in different zones. Hence, the persistence and weighted mean analogue forecasts had lower forecasting performances compared to the variation analogue forecast, particularly for the high-response catchment of Sirikit Dam. The variation analogue forecast also had a better forecasting performance than the ANN and WANN models. It was also superior to the other forecasting methods when forecasting extreme values. ACKNOWLEDGMENTS: This study was financially supported by the Science and Technology Research Partnership for Sustainable Development (SATREPS). We also thank the Agricultural Research Development Agency (ARDA) of Thailand for providing a research grant to the first author. This work was supported by the Program for Risk Information on Climate Change from the Ministry of Education, Culture, Sports, Science, and Technology-Japan (MEXT). This research was also supported by the Environment Research and Technology Development Fund (S-) of the Ministry of the Environment, Japan. REFERENCES. Mohammadi, K., Esmali, H.R. and Dardashti, Sh.D.: Comparison of regression, ARIMA and ANN models for reservoir forecasting using snowmelt equivalent (a case study of Karaj), J. Agric. Sci. Technol., Vol.7, pp.7, 5.. Zhang, J., Cheng, C., Liao, S., Wu, X. and Shen, J.: Daily reservoir forecasting combining QPF into ANNs model, Hydrol. Earth Syst. Sci. Discuss., 6(), pp. 5, 9.. Muluye, G.Y. and Coulibaly, P.: Seasonal reservoir forecasting with low-frequency climatic indices: a comparison of data driven methods, Hydrological Sciences Journal, 5:, pp.58 5, DOI:.6/hysj Budu, K.: Comparison of Wavelet-Based ANN and Regression Models for Reservoir Inflow Forecasting., J. Hydrol. Eng., 9(7), pp Svensson, C.: Seasonal river flow forecasts for the United Kingdom using persistence and historical analogues, Hydrological Sciences Journal, DOI:.8/ , Bridhikitti, A.: Connection of ENSO/IOD and aerosols with Thai rainfall anomalies and associated implications for local rainfall forecasts, Int. J. Climatol., Vol., pp ,. 7. Amnatsan, S., Kuribayashi, D. and Jayawardena, A.W.: Application of artificial neural networks and wavelet analysis in prediction of water level in Nan River of Thailand, Proceeding of Annual Conference rd Annual Conference, Japan Society of Hydrology and Water Resources.,. (Received September, 5)