Utilizing Artificial Intelligence Techniques for a Long–Term Water Resource Assessment in the ShihMen Reservoir for Water Resource Allocation

Abstract

Accurate long–term water resource supply simulation and demand estimation are crucial for effective water resource allocation. This study proposes advanced artificial intelligence (AI)–based models for both long–term water resource supply simulation and demand estimation, specifically focusing on the ShihMen Reservoir in Taiwan. A Long Short–Term Memory (LSTM) network model was developed to simulate daily reservoir inflow. The climate factors from the Taiwan Central Weather Bureau’s one–tiered atmosphere–ocean coupled climate forecast system (TCWB1T1) were downscaled using the K–Nearest Neighbors (KNN) method and integrated with the reservoir inflow model to forecast inflow six months ahead. Additionally, Multilayer Perceptron (MLP) and Gated Recurrent Unit (GRU) were employed to estimate agricultural and public water demand, integrating both hydrological and socio–economic factors. The models were trained and validated using historical data, with the LSTM model demonstrating a strong ability to capture seasonal variations in inflow patterns and the MLP and GRU models effectively estimating water demand. The results highlight the models’ high accuracy and robustness, offering valuable insights into regional water resource allocation. This research provides a framework for integrating AI–driven models with Decision Support Systems (DSSs) to enhance water resource management, especially in regions vulnerable to climatic variability.

1. Introduction

Accurate long–term water resource supply and demand assessments are essential for effective water resource allocation [1,2,3,4,5]. Given the increasing variability in climatic conditions and socio–economic activities, these simulations and estimations become even more critical for sustainable water management. Long–term forecasts are crucial for planning and managing water resources, especially in regions where water availability is influenced by seasonal variations and extreme weather events. As highlighted by several studies, effective water allocation relies heavily on precise forecasting of both supply and demand. For instance, ref. [6] highlights the impact of projected climate change and human activities on the optimal allocation of water resources, emphasizing the importance of considering both environmental and anthropogenic factors and identifying water availability and demand as prerequisites for optimal water resource management. Similarly, [5] addresses the critical issue of the water resource supply–demand imbalance, emphasizing the importance of water resource allocation (WRA) models based on supply–demand forecasts and comprehensive value assessments for sustainable development.

In the context of Taiwan, where this study is focused, the challenges are further compounded by the frequent occurrence of typhoons and the subsequent dry periods. These climatic characteristics necessitate a highly adaptive and responsive water resource management system that can balance the dual needs of flood control and drought preparedness. The integration of advanced forecasting techniques into water resource management practices is therefore not just beneficial but essential. Chang et al. [7] underscore the critical importance of accurately forecasting both agricultural and public water use, particularly during drought periods, to ensure that water allocation decisions effectively balance the needs of different sectors and mitigate the potential impacts on food supply and public health. Kim et al. [8] indicate that to effectively manage the allocation of limited water resources among agricultural, municipal, and environmental uses, tools like the Water Evaluation and Planning (WEAP) system are essential. The WEAP integrates supply, demand, water quality, and ecological considerations into a practical tool for comprehensive water resource planning, ensuring that all factors are considered in decision–making processes.

Understanding and accurately forecasting water supply and demand are crucial for effective water resource management and allocation. This necessity has led to an increasing focus on applying machine learning techniques in water resource forecasting. Machine learning models have shown great promise in predicting both water supply and demand with higher accuracy compared to traditional statistical methods [3,4,9,10,11,12,13,14,15,16,17,18,19]. The ability of Long Short–Term Memory (LSTM) networks to learn long–term dependencies between input and output variables makes it particularly effective for capturing storage effects in catchments influenced by snow or other complex hydrological processes [20]. Kratzert et al. [21] demonstrated that LSTM networks outperform traditional hydrological models, like the Sacramento Soil Moisture Accounting Model (SAC–SMA). Using data from multiple catchments, LSTM models can improve performance when applied to individual catchments, underscoring their versatility and robustness in hydrological applications. Similarly, Multilayer Perceptrons (MLPs) have been proven effective in forecasting short–term urban water demand. A study conducted in Guaratuba, Brazil found that MLP outperformed other machine learning models in predicting daily water demand using historical, meteorological, and calendar data. This research highlighted the complexity of water demand forecasting and showed that MLP, when validated with appropriate techniques, provides accurate predictions that are crucial for efficient water resource management [22]. Additionally, Gated Recurrent Unit (GRU) networks have been shown to outperform conventional artificial neural networks in short–term water demand forecasting. Guo and Liu [23] developed a GRU–based model to forecast water demand at 15–min intervals up to 24 h ahead, demonstrating superior performance over traditional ANN models. These results emphasize the flexibility and effectiveness of GRU models in handling the non–linear nature of water demand changes, making them valuable tools for water resource management. These models’ capabilities to process large datasets and continuously learn from new data make them powerful tools for improving the precision of water resource management practices. By leveraging the strengths of LSTM, MLP, and GRU models, water managers can better anticipate future water needs and optimize resource distribution to meet the demands of both the agricultural and public sectors.

While the existing literature has primarily focused on short–term predictions based on observed data, these short–term forecasts offer limited utility when it comes to the long–term allocation of water resources in reservoir catchment areas, particularly for managing the future needs of the public and agricultural sectors. Short–term fluctuations, while valuable for immediate operational adjustments, do not provide the comprehensive outlook required for effective planning, which must consider future reservoir inflows and the anticipated demand of both sectors. This foresight is essential to ensure a balanced and sustainable distribution of water resources over extended periods. Recognizing this limitation, the present study seeks to bridge the gap by integrating the strengths of machine learning techniques with the predictive capabilities of numerical weather models, specifically the TCWB1T1 model developed by the Taiwan Central Weather Bureau. The TCWB1T1 is a one–tiered atmosphere–ocean coupled climate forecast system that provides climate simulations for up to six months ahead, offering critical data on temperature and precipitation [24]. By extending the forecasting horizon and incorporating these climate simulations, this research addresses the critical need for long–term water resource management strategies. Utilizing the simulated results from TCWB1T1 for the next six months, this approach aims to generate long–term forecasts for both water supply and demand, enabling water resource managers to make informed decisions that optimize the allocation of resources based on anticipated future conditions.

To address these challenges and improve the efficacy of water resource management in Taiwan, this study proposes a novel approach that integrates advanced machine learning models with numerical weather predictions. By leveraging the predictive power of LSTM, MLP, and GRU models, combined with climate simulations that predict data for up to six months ahead, this study aims to develop a comprehensive framework for long–term water supply and demand forecasting. This integrated approach enhances the accuracy of water resource predictions and provides a robust tool for managing the complexities associated with reservoir operations, particularly in the face of Taiwan’s unique climatic conditions. The following sections will detail the methodology employed, the data sources utilized, and the validation processes implemented to ensure the reliability of the proposed forecasting model.

2. Materials and Methods

2.1. Study Area

The ShihMen Reservoir is a crucial multi–purpose reservoir in Northern Taiwan, with a current storage capacity of approximately 199 million m³ (designated storage capacity of 309 million m³). It plays a vital role in serving the agricultural irrigation, domestic water supply, and industrial water needs of Taoyuan City, New Taipei City, and Hsinchu County. Additionally, the reservoir is entrusted with multiple responsibilities, including flood control, power generation, tourism promotion, and ecological environment preservation.

The ShihMen Reservoir is located in the middle reaches of the DaHan River, with the terrain sloping from south to north. The reservoir’s catchment area spans approximately 763.4 km². The region experiences an annual average rainfall of about 2600 mm, with more than 60% of the total precipitation occurring during the wet season from May to October, primarily due to heavy rain associated with typhoons. Figure 1 illustrates the location of the ShihMen Reservoir catchment.

2.2. Methodology Construction

To assist in creating a water resource allocation system (WRAS), this study proposes a long–term water resource supply simulation model and a long–term water resource demand estimation model. Figure 2 shows a flowchart of the proposed models. The reservoir inflow model is constructed with a daily temporal resolution, aligning with the characteristics of the catchment area to capture the variability in inflow patterns accurately. However, for the practical application of water resource management and allocation, the forecasts are aggregated on a ten–day basis, which is the standard time frame used for decision making in water resource scheduling.

The long–term forecasts generated by the model provide forecasts for 18 ten–day periods (six months) for reservoir inflow, agricultural water demand, and public water demand. This approach ensures that the forecasts are both accurate and usable for stakeholders, balancing the need for detailed, high–resolution data with the practical requirements of water resource management. Details of the proposed models are described below.

2.2.1. Simulation Model of Long–Term Water Resource Supply

An illustration of the proposed long–term water resource supply simulation model is shown in Figure 3. Initially, a reservoir inflow simulation model is constructed using historical observational data. The computation kernel of the model adopts the LSTM method, with the observed rainfall and temperature as input variables and reservoir inflow as the output variable. LSTM networks, known for their ability to capture long–term dependencies in sequential data, are particularly suitable for modeling hydrological processes due to their capacity to handle complex temporal patterns and non–linear relationships in time series data [21,24,25]. The LSTM–based model leverages historical inflow, rainfall, and temperature data, ensuring that the simulation outputs accurately capture the unique seasonal variations in Taiwan’s climate.

Subsequently, to simulate the long–term variation in reservoir inflow, data from the Taiwan Central Weather Bureau’s one–tiered atmosphere–ocean coupled climate forecast system version 1 (TCWB1T1) is integrated, serving as input for future scenario simulations. TCWB1T1 is a dynamic statistical climate forecasting system developed for short–term climate predictions [26]. The atmospheric model utilizes the CWB Global Atmosphere Model developed by the Central Weather Bureau [27], while the ocean model employs the Geophysical Fluid Dynamics Laboratory (GFDL) Modular Ocean Model version 3 (MOM3) [28]. These two models are coupled once a day to produce data for the next six months. However, due to the relatively large horizontal (approximately 100 km) resolution of the TCWB1T1 model, downscaling methods are required to make it applicable to Taiwan’s catchment–scale regions. Downscaling methods can be broadly categorized into dynamical downscaling and statistical downscaling. Statistical downscaling uses observed data from weather stations to establish a statistical relationship with the results of climate models, enabling the conversion of future climate simulations into catchment–scale station data for model inputs. This approach is advantageous over dynamical downscaling as it is generally less computationally intensive and can be more easily tailored to local climate characteristics. The K–Nearest Neighbors (KNN) method, being a non–parametric statistical approach, has been widely applied to downscaling problems due to its convenience [29,30,31]. The K–Nearest Neighbors (KNN) method is employed to build downscaling models for rainfall and temperature separately. Taking the rainfall downscaling model as an example, the KNN method is used to establish the relationship between historical coarse–grid rainfall data and ground–level observed rainfall. Once constructed, this downscaling model can be used to generate catchment–scale rainfall data by using the future 6–month rainfall simulation results from TCWB1T1.

Finally, based on the downscaled rainfall and temperature data for the next six months, these data are input into the previously established reservoir inflow simulation model. The daily reservoir inflow for the next six months can then be generated, providing information on the long–term water resource supply as a reference for water resource allocation.

While the current model primarily incorporates rainfall and temperature factors, it is acknowledged that additional variables such as land use changes, soil moisture, and groundwater levels could further enhance the model’s comprehensiveness. However, since the TCWB1T1 model used for input data only provides forecasts for temperature and rainfall for the next six months, incorporating these additional variables would require separate models to forecast each of these factors, introducing additional uncertainties that could affect the overall reliability of the proposed model. As a result, this study focuses on utilizing the factors provided by the TCWB1T1 model for constructing the long–term water resource supply simulation model, ensuring a balance between model accuracy and the practical availability of forecast data.

2.2.2. Long–Term Water Resource Demand Assessment Model

An illustration of the proposed long–term water resource demand estimation model is shown in Figure 4. To assess the future 6–month water demand for both public and agricultural sectors, it is necessary to collect factors related to water demand and establish a water demand estimation model. This is similar to the approach used in Korea’s Long–term Comprehensive Water Resources Plan, which categorizes water demand into high, medium, and low scenarios based on domestic, industrial, and irrigation uses over a 10–year planning period. Socio–economic trends and sector–specific demand factors should be considered to estimate future water needs accurately [8]. As noted in the literature, three major categories of variables—consumption, climatic, and socio–economic—are typically used as predictors in soft computing methods for water demand forecasting. Consumption variables include historical patterns, and climatic variables account for parameters such as temperature and rainfall, while socio–economic variables consider factors like population growth and economic conditions [9].

For public water demand, the following data are collected: 10–day average rainfall ($R_{10–\mathrm{d}\mathrm{a}\mathrm{y}})$, 10–day average temperature at Daxi station ($T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}})$, 10–day average temperature at Fuxing station ($T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{F}\mathrm{u}\mathrm{x}\mathrm{i}\mathrm{n}\mathrm{g}})$, 10–day average reservoir inflow ($I_{10–\mathrm{d}\mathrm{a}\mathrm{y}})$, planned public water usage ($V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d}\mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}})$, number of employees ($n_{\mathrm{e}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{o}\mathrm{y}\mathrm{e}\mathrm{e}})$, GDP, economic growth rate (quarterly SAQR comparison) ($GDP)$, economic growth rate (year–on–year (YoY) comparison) ($EG)$, number of factories ($n_{\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{y}})$, population ($n_{\mathrm{p}\mathrm{o}\mathrm{p}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}})$, historical 10–day average public water supply ($V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}})$, and the ranking values of historical 10–day public water supply ($R_{\mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}})$. To estimate future public water demand, a GRU model is employed as the computational core, learning the non–linear relationships between these factors and the 10–day average system water demand. When estimating the future 6–month public water demand, the projected values of selected input factors are input into the GRU–based public water demand estimation model to generate the anticipated public water demand for the next six months.

The agricultural water demand estimation model is constructed using a similar approach. For agricultural water demand, the following data are collected: $R_{10–day}$, $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{F}\mathrm{u}\mathrm{x}\mathrm{i}\mathrm{n}\mathrm{g}}$, $I_{10–day}$, rice field area ($A$), planned agricultural water usage ${(V}_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d}\mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}})$, historical 10–day average agricultural water supply ($V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}})$, and the ranking values of historical 10–day agricultural water supply ($R_{\mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}})$. The MLP model is trained to capture the complex, non–linear relationships between these factors and the 10–day average system water demand. The projected values of selected input factors are then input into the MLP–based agricultural water demand estimation model to forecast the anticipated agricultural water demand for the next six months.

2.2.3. Model Evaluation

To evaluate the model’s performance, four performance indices are employed to evaluate the discrepancy between the observed and forecasted/simulated/estimated values: the root mean square error (RMSE), mean absolute error (MAE), coefficient of correlation (CC), and coefficient of efficiency (CE). The coefficient of correlation (CC) is closely related to the coefficient of determination (R²), as R² is the square of CC, indicating the proportion of variance in the dependent variable that is predictable from the independent variables. The coefficient of efficiency (CE) is equivalent to the Nash–Sutcliffe efficiency (NSE), which is commonly used in hydrological modeling to evaluate forecasting ability. Smaller RMSE and MAE values indicate less significant errors between the observed and forecasted values, whereas a higher CC value means better agreement between the observed and forecasted/simulated/estimated values. The CE value is used to evaluate forecasting ability, with a CE value equal to 1 representing perfect performance.

3. Results and Discussion

3.1. The Performance of the Long–Term Water Resource Supply Simulation

Initially, an LSTM–based daily inflow simulation model was constructed using historical rainfall and temperature data. Considering the significant variations in Taiwan’s water conditions across different months, the daily flow model was categorized into four models based on climatic characteristics: LSTM–11T1, LSTM–2T4, LSTM–5T6, and LSTM–7T10. These models correspond to the data from November to January, February to April, May to June, and July to October, respectively. The division between training and testing data was carried out on an annual basis. Data from the years 2011, 2015, 2017, and 2019 were randomly selected as the testing data, while the remaining data were used for model training. A fixed training–to–testing data ratio of 11:5 was maintained to ensure consistency in evaluating the model’s performance. While cross–validation could potentially provide additional insights, it often results in varying optimal input factor combinations across different runs, which can lead to uncertainty in determining the most appropriate input factors for practical application. Therefore, a fixed data ratio was used to provide clear and reliable results in this study. The LSTM–based daily inflow simulation model can be written in a general form as follows:

$$Q_d=f_{\mathit{daily}\mathit{I}}(R_d,R_{d-1},{\cdots ,R}_{d-(L_R-1)},T_d)$$

(1)

where d is the current day; Q_d, R_d, and T_d are the inflow, rainfall, and temperature at day d, respectively; and L_R denotes the lag length of rainfall. The parameter search ranges and optimal combinations are shown in

Table 1. Parameter search ranges and optimal combinations of the simulation model.

Parameter	Search Ranges	LSTM–11T1	LSTM–2T4	LSTM–5T6	LSTM–7T10
$L_R$	20~30	21	21	21	21
Hidden layer	{(16), (32)} {(32), (64)} {(64), (128)}	{(16), (32)}	{(64), (128)}	{(16), (32)}	{(64), (128)}
Activation	Tanh Relu	Relu	Relu	Relu	Relu
Optimizer	Nadam Adam Rmseprop Lbfgs	Rmsprop	Rmsprop	Rmsprop	Rmsprop
Batch Size	8 16 32	16	16	16	16
Dropout	10% 5% 0%	0%	0%	0%	0%
Epoch	100 150	150	100	150	150

. Based on the grid search method, the optimal combinations selected for different LSTM models include the lag length, hidden layers, activation functions, optimizers, batch sizes, dropout rates, and epochs. For instance, the LSTM–11T1 and LSTM–5T6 models use a lag length of 21, {(16), (32)} hidden layers, a Relu activation function, Rmsprop optimizer, a batch size of 16, a 0% dropout ratio, and 150 epochs.

It is important to note that while the inflow simulation model used in this study is constructed with machine learning techniques, which generally have less dependence on specific terrain conditions compared to traditional hydrological models, the model still requires rigorous training and testing to ensure good generalization performance. This is particularly crucial when applying the model to different terrains or environmental conditions, as variations in these factors could impact the model’s accuracy.

Figure 5 presents the LSTM–based daily inflow simulation model results. As shown in Figure 5, the simulated inflow yielded by the proposed model is in good agreement with the observed data. The performance indices of the LSTM–based daily inflow simulation are presented in

Table 2. The RMSE, MAE, CE, and CC of the training and testing data for LSTM–based models.

Model		RMSE (cms)	MAE (cms)	CE	CC
LSTM–11T1	Training	4.0	3.1	0.65	0.87
	Testing	10.6	6.4	0.65	0.87
LSTM–2T4	Training	5.2	4.2	0.90	0.97
	Testing	8.0	5.0	0.46	0.86
LSTM–5T6	Training	12.1	8.3	0.70	0.90
	Testing	24.2	13.2	0.73	0.88
LSTM–7T10	Training	28.8	12.5	0.90	0.96
	Testing	33.6	16.1	0.88	0.95

. Due to frequent typhoon impacts resulting in heavy rainfall and high inflow from July to October in Taiwan, the LSTM–7T10 simulation exhibits slightly larger errors than other models. Nevertheless, it maintains a correlation coefficient above 0.95, indicating that the proposed model can effectively simulate daily reservoir inflow based on rainfall and temperature data.

Next, a KNN–based downscaled model was developed to downscale the regional large–scale daily rainfall and temperature data produced by the TCWB1T1 model to the catchment meteorological stations. The KNN model calculates the similarity between the historical data and the target day’s meteorological factors using the Euclidean distance, where shorter distances indicate more significant similarity. After sorting, the model selects the average values of the most similar days for calculation. The calculation formula is as follows:

$$d_T^n=\sqrt{{\sum }_{i=1}^m{\left(F_i^n-f_{i,T}^n\right)}^2}$$

(2)

where n is the data years, T is the search month of the moving window, m is the total number of meteorological factors, and F and f denote the target day and historical data, respectively.

Figure 6 presents the KNN–based downscaling model results. As shown in Figure 6, the 10–day rainfall and average temperature simulated by the proposed model closely match the observed data. The statistical characteristics of both the observed data and simulation results are summarized in

Table 3. Statistical characteristics of observed data and simulation results.

Factor		Mean	Standard Deviation	Skewness Coefficient	Rainfall Probability
10–day average rainfall (ShihMen station)	Observed	7.3	21.0	6.4	0.36
	Simulated	7.2	16.7	6.1	0.55
10–day average rainfall (YuFeng station)	Observed	5.3	20.4	10.7	0.32
	Simulated	5.6	20.0	10.1	0.47
10–day average temperature (FuXing station)	Observed	20.1	4.6	−0.3	–
	Simulated	20.1	4.5	−0.3	–
10–day average temperature (DaSi station)	Observed	21.6	5.1	−0.2	–
	Simulated	21.6	5.0	−0.2	–

. The table presents the statistical characteristics of the observed and simulated values, demonstrating the downscaling model’s ability to accurately convert the large–scale temperature factor from the TCWB1T1 model output to the catchment scale. However, the standard deviation of the large–scale rainfall factor after downscaling exhibits a slightly larger difference at the ShihMen station, which is attributed to the inherent variability in rainfall data. Nevertheless, the simulated mean values closely align with the observed values. This suggests that, when considered on a 10–day basis, the simulated average rainfall is in close agreement with the observed values. These downscaled results can serve as input for the previously established LSTM–based inflow simulation model.

Since Taiwan’s dry season is concentrated in the winter and spring, forecasting the water supply for a specific time is crucial for long–term water resource management. In this study, the future 6–month temperature and rainfall output results from TCWB1T1, obtained on 30 September 2019, are incorporated. Using the proposed KNN–based downscaling model, the data are transformed to the catchment scale. Subsequently, these downscaled data are input into the proposed LSTM–based daily inflow simulation model to forecast the reservoir inflow for the next six months (from November to March).

Figure 7 illustrates a comparison between the observed inflow and the long–term forecasts generated by the LSTM–based inflow simulation model on 30 September 2019. The forecasted inflow for the first ten days of October appears to underestimate the observed inflow. This discrepancy is attributed to Typhoon MITAG, which affected Taiwan on 30 September 2019. Although the typhoon gradually moved away from Taiwan on 1 October, its outer circulation continued to bring significant rainfall to the northern regions of Taiwan. This resulted in the actual inflow being higher than forecasted. The output results of climate factors from the TCWB1T1 model, when processed through the proposed downscaling model, fail to capture the rainfall and reservoir inflow associated with the typhoon, resulting in an underestimation. Nevertheless, the forecasting results for the long–term reservoir inflow in other periods are close to the observed values, serving as a reference for future long–term water resource supply.

Based on the evaluation results, the LSTM–based model demonstrates a strong ability to capture seasonal variations in inflow patterns, as evidenced by high correlation coefficients and low error margins. However, while the model provides accurate long–term inflow forecasts, there remains inherent uncertainty in these predictions, particularly during extreme weather events like typhoons. Given this uncertainty, current reservoir management practices, especially for large reservoirs like ShihMen, continue to rely primarily on established operation rule curves and real–time adjustments based on actual water conditions. When the reservoir’s effective storage capacity is between the upper and lower rule curves, water supply is provided up to the planned demand. However, when the effective storage capacity falls between the lower rule curves and the critical lower rule curves, public water supply is reduced by 10%, and agricultural water supply is reduced by 25% to conserve water resources [7]. This method allows reservoir authorities to manage water supplies more effectively, ensuring that critical demands are met while maintaining long–term sustainability.

The proposed model is focused on long–term forecasts, up to six months ahead, to manage water resource allocation better. When future predictions indicate an ample water supply and high reservoir water levels, the model can help optimize water usage, allowing for more efficient resource management. Conversely, suppose the model forecasts unfavorable water conditions. In that case, it may be necessary to consider implementing water–saving measures to meet future demand, particularly if there is no indication of improvement in water availability. This can be particularly beneficial for optimizing water resource allocation across different sectors, such as agricultural and public use, during different seasons.

Regarding extreme weather events like typhoons, forecasting such events over long–term periods presents significant challenges with current technology. Therefore, it is recommended that separate numerical weather models be utilized to forecast extreme weather as it approaches. These short–term forecasts can then be integrated with flood management operations to ensure that both long–term water resource planning and immediate operational adjustments are effectively managed.

The long–term forecasts produced by the model are intended to serve as supplementary information, offering valuable background data to inform decision making. These forecasts are designed to support, rather than replace, the dynamic management strategies employed by reservoir authorities, providing a broader context for anticipating future scenarios and optimizing water resource allocation. This approach ensures that while the model contributes important insights into potential future inflow conditions, it also acknowledges the critical role of real–time data and adaptive management in maintaining effective reservoir operations. As highlighted by Kim et al. [12], effective adaptive management of water demand involves driving near–future information through an iterative learning process amidst uncertainty. This underscores the importance of providing accurate, timely information to inform and implement flexible, responsive policies. Integrating these forecasts into the decision–making process can enhance the overall resilience and effectiveness of water resource management strategies.

3.2. The Performance of the Long–Term Water Resource Demand Estimation

Estimating long–term water resource demand is a challenging task as it involves various factors related to both hydrological and socio–economic environments. In Taiwan, agricultural water use is primarily managed by the Ministry of Agriculture, while other types of water use, including domestic water supply and industrial water use, are managed by the Taiwan Water Corporation. Therefore, historical agricultural and public water supply volumes can be considered as indicators of water demand, and regression relationships can be established using artificial intelligence methods with hydrological and socio–economic factors.

Agricultural water usage data are available for a longer duration. In this study, data spanning from 2003 to 2019, a total of 17 years, were collected. Data from the years 2007, 2011, 2015, 2017, and 2019 were chosen as the testing dataset, while the remaining data were utilized as the training dataset. The MLP model was employed to estimate agricultural water demand utilizing optimal input combinations and parameters determined through a grid search method. The parameter search ranges and optimal combinations are shown in

Table 4. The parameter search ranges and optimal combinations of the agricultural estimation model.

Parameter	Search Ranges	MLP
Input	$R_{10–\mathrm{d}\mathrm{a}\mathrm{y}}$, $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{F}\mathrm{u}\mathrm{x}\mathrm{i}\mathrm{n}\mathrm{g}}$, $I_{10–\mathrm{d}\mathrm{a}\mathrm{y}}$, $A$, $V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$, $V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$, $R_{\mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$	$T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $A$, $V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$, $V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$
Hidden layer	{(16)} {(32)} {(64)} {(128)}	{(64)}
Activation	Tanh Relu	Relu
Optimizer	Nadam Adam Rmseprop Lbfgs	Lbfgs
Batch Size	8 16 32	16
Learning rate	0.001 0.0001 0.00001	0.0001

. The grid search identified the optimal configuration for the MLP model, which included a hidden layer size of 64, a Relu activation function, an Lbfgs optimizer, a batch size of 16, and a learning rate of 0.0001. The optimal input combination for the model included $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $A$, $V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$, and $V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{a}\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$.

From the selected factors, it can be inferred that the size of the cultivated area is directly related to water demand. The temperature factor influences the timing of rice planting, which indirectly reflects water needs at different times. The inclusion of planned and historical water usage allows the model to learn from the region’s historical water usage patterns and, when combined with planned water usage, predict future demand. Interestingly, rainfall and reservoir inflow were not selected as input factors, indicating that farmers’ decisions regarding rice cultivation are not directly influenced by the availability of water resources. This suggests that agricultural practices may not be sufficiently aligned with actual water supply conditions. Therefore, if the current reservoir levels are low and there is a risk of drought, it would be prudent to inform agricultural authorities about the potential for future water restrictions. Early notification could help these stakeholders prepare and adapt their plans accordingly, ensuring a more efficient approach to water resource allocation.

Figure 8 illustrates the outcomes of the MLP–based agricultural water demand model. The results indicate that due to unfavorable water conditions in the Taoyuan area during the first half of 2015, characterized by a low reservoir storage level, the government implemented fallow land policies. As a result, the actual agricultural water usage (represented by the blue dots) during that period was significantly lower than in other years. The proposed model also estimated lower agricultural water demand during that period. The performance indices of the MLP–based estimation model are presented in

Table 5. The RMSE, MAE, CE, and CC of the training and testing data for the MLP–based model.

Model		RMSE (10,000 m3)	MAE (10,000 m3)	CE	CC
MLP	Training	27.7	22.4	0.64	0.91
	Testing	32.1	24.3	0.50	0.88

. The results demonstrate that the proposed model attains a high correlation coefficient (CC) value of 0.88 on the testing dataset, indicating a significant level of accuracy in the estimation results.

Data related to public water supply were obtained for the years 2003 to 2019, spanning a total of 17 years, with the data from 2011, 2015, 2017, and 2019 being designated as the testing dataset, while the remaining data were used for the training dataset. The GRU model was employed to estimate public water demand utilizing optimal input combinations and parameters determined through a grid search method. The parameter search ranges and optimal combinations are shown in

Table 6. The parameter search ranges and optimal combinations of the public estimation model.

Parameter	Search Ranges	GRU
Input	$R_{10–\mathrm{d}\mathrm{a}\mathrm{y}}$, $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{F}\mathrm{u}\mathrm{x}\mathrm{i}\mathrm{n}\mathrm{g}}$, $I_{10–\mathrm{d}\mathrm{a}\mathrm{y}}$, $V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$, $n_{\mathrm{e}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{o}\mathrm{y}\mathrm{e}\mathrm{e}}$, GDP, $GDP$, $EG$, $n_{\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{y}}$, $n_{\mathrm{p}\mathrm{o}\mathrm{p}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}$, $V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$, $R_{\mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$	$T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $n_{\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{y}}$, $V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$, $V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$
Hidden layer	{(16), (32)} {(32), (64)} {(64), (128)}	{(16), (32)}
Activation	Tanh Relu	Relu
Optimizer	Nadam Adam Rmseprop Lbfgs	Nadam
Batch Size	8 16 32	8
Dropout	10% 5% 0%	5%
Epoch	100 150	150

. The grid search identified the optimal configuration for the GRU model, which included a hidden layer size of {(16), (32)}, a Relu activation function, a Nadam optimizer, a batch size of 8, a 5% dropout ratio, and 150 epochs. The optimal input combination for the model included $T_{10–\mathrm{d}\mathrm{a}\mathrm{y},\mathrm{D}\mathrm{a}\mathrm{x}\mathrm{i}}$, $n_{\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{y}}$, $V_{\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{n}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$, $V_{\mathrm{h}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{p}\mathrm{u}\mathrm{b}\mathrm{l}\mathrm{i}\mathrm{c}}$.

The selection of these factors shows that public water demand is closely tied to the number of operational factories, reflecting the direct relationship between industrial activity and water consumption. The inclusion of temperature as an input factor may capture seasonal variations in industrial operations or other temperature–dependent processes. Meanwhile, using planned and historical water usage data allows the model to understand the region’s past and anticipated water use patterns, thereby improving the accuracy of future demand predictions. Notably, rainfall and reservoir inflow were also not selected as input factors, indicating that the availability of water resources does not directly influence industrial and domestic water demand in the region. This observation aligns with findings from the Han River basin in Korea, where studies have shown that public water use remains relatively stable across different months regardless of runoff variations [8]. This suggests that regional industrial activities and domestic water needs continue to be uninterrupted despite fluctuations in water supply, as these sectors are less impacted by changes in water availability due to the management authorities’ obligation to meet basic requirements. As a result, during periods of low water availability, management authorities face the critical challenge of ensuring a consistent supply of essential public water. This highlights the importance of effective water resource management, especially during periods of scarcity, to balance agricultural and public water use.

Figure 9 illustrates the outcomes of the GRU–based public water demand model. The performance indices of the GRU–based estimation model are presented in

Table 7. The RMSE, MAE, CE, and CC of the training and testing data for the GRU–based model.

Model		RMSE (10,000 m3)	MAE (10,000 m3)	CE	CC
GRU	Training	4.8	3.9	0.40	0.71
	Testing	4.8	4.0	−0.14	0.57

. The results indicate that the estimates for 2011 and 2019 closely align with the actual public water supply, while disparities exist for 2015 and 2017. These disparities are attributed to unfavorable water conditions in 2015, leading to multiple measures implemented by government authorities, including water pressure reduction and water restrictions. In March 2017, the first stage of water restrictions was also implemented. These measures, which influenced public water consumption due to human–made decisions, resulted in a significant difference between the model results and actual values for those years. Nevertheless, with the growth in population and industry in the Taoyuan area, the proposed model also indicates an increasing trend in public water demand and can provide valuable references for water resource management by estimating potential public water demand based on hydrological and socio–economic data characteristics.

3.3. Integration with Decision Support Systems (DSSs)

To effectively utilize the proposed long–term water resource supply simulation and demand estimation models in real–world applications, integration with Decision Support Systems (DSSs) is crucial. The models developed in this study are designed to be compatible with existing DSS frameworks used by reservoir management authorities. The integration process involves the following key steps:

• Data Integration: The DSS interface feeds real–time data from meteorological stations and relevant sources into the models.
• Model Execution: The DSS triggers the models to run using the latest data inputs, generating updated forecasts.
• Output Interpretation: The DSS interprets the model outputs, including predictions of inflow and water demand, to inform water management strategies.
• Decision–Making Support: The DSS utilizes the model outputs to suggest or automate decisions, such as reservoir release schedules or water rationing measures.

While this study demonstrates the potential of these models to be integrated with DSSs, future research could focus on implementing and testing this integration in operational environments. This will help validate the practical utility of the models and refine their application in DSSs for better decision making in water resource management.

4. Conclusions

Accurate forecasts for long–term supply and demand are crucial in water resource allocation. For this purpose, this study proposes artificial intelligence–based models for long–term water resource supply simulation and long–term water resource demand estimation.

This study focuses on the ShihMen Reservoir in Taoyuan. It employs LSTM based on historical observed rainfall and temperature data as input and reservoir inflow as output to develop a daily inflow simulation model. The results indicate that the constructed model (LSTM–7T10) exhibits larger errors from July to October due to significant variations in reservoir inflow. However, the correlation coefficient remains at 0.95, demonstrating the model’s ability to capture the overall trends despite the variations. In addition, the rainfall and temperature outputs from the Taiwan Central Weather Bureau TCWB1T1 for the next six months are integrated. By employing a KNN–based downscaling model, these factors are downscaled to the catchment scale, providing rainfall and temperature data for the next six months. These downscaled data are input into the daily inflow simulation model to generate reservoir inflow for the next six months. The results indicate that the downscaled model output for rainfall and temperature closely matches the observed data. Moreover, the future 6–month flow generated by the daily inflow simulation model aligns with the actual inflow trends, serving as a valuable reference for long–term water resource supply.

The future long–term water resource demand is further divided into public water use and agricultural water use. For each category, the MLP and GRU models are employed to develop estimation models for agricultural and public water demand, respectively. In this study, the actual volumes of agricultural and public water supply are treated as indicators of water demand. For agricultural water demand, the optimal input combination underscores that the size of cultivated areas and temperature are directly related to agricultural water demand, with temperature influencing the timing of rice planting and thereby reflecting seasonal water needs. The inclusion of planned and historical agricultural water usage data in the model allows us to learn from past usage patterns to forecast future demand. For agricultural water demand, the selection of input factors demonstrates that public water demand is closely linked to the number of operational factories, reflecting the direct relationship between industrial activity and water consumption. The inclusion of temperature as an input factor may account for seasonal variations in industrial operations or other temperature–dependent processes. Additionally, using planned and historical water usage data enables the model to recognize the region’s past and anticipated water use patterns, thus enhancing the accuracy of future demand predictions. Notably, rainfall and reservoir inflow were not selected as input factors, suggesting that the availability of water resources does not directly influence both agricultural and public water demand in the region. This illustrates that regional industrial activities and domestic water needs are sustained even during water scarcity, likely due to the government’s commitment to maintaining a stable water supply for essential domestic and industrial needs.

Despite the strong performance of the models, this study acknowledges the inherent uncertainties in long–term forecasting, particularly under extreme weather conditions. As such, these AI–based models are intended to supplement existing management practices by providing additional data to inform decision making rather than replace established operational strategies. The findings underscore the potential of integrating AI techniques with traditional reservoir management practices to improve the overall efficiency of water resource allocation, especially in regions prone to climatic variability. Further research and practical implementation are encouraged to validate these models in operational environments and refine their application in real–world decision–making processes.

5. Future Recommendations

This study has demonstrated the potential of AI–driven approaches to improve long–term water resource forecasting, particularly by applying the LSTM, MLP, and GRU models. However, several areas warrant further investigation to enhance the applicability and reliability of these models under varying environmental and climatic conditions:

A.

Cross–Validation Techniques: While this study employed a fixed training–to–testing data ratio to ensure consistency in evaluating model performance, future research could benefit from incorporating cross–validation techniques. Cross–validation can provide a more robust assessment by mitigating overfitting or underfitting, although it may result in varying optimal input factor combinations across different runs. Exploring this variability’s impact and establishing best practices for input factor selection could maintain model reliability and applicability in diverse contexts.

B.

Integration of Additional Variables: The current models primarily incorporate rainfall, temperature, and socio–economic factors. Future studies should consider integrating additional variables such as land use changes, soil moisture, and groundwater levels. However, including these factors would necessitate the development of separate predictive models, which could introduce additional complexities and uncertainties. Managing these challenges effectively ensures model robustness and accuracy in long–term forecasts.

C.

Forecasting Extreme Weather Events: Given the challenges in accurately forecasting extreme weather events like typhoons over long–term periods, future research should explore integrating short–term numerical weather models designed for such forecasts. A more comprehensive and adaptive water resource management strategy can be developed by combining these short–term predictions with long–term inflow forecasts. This approach would enhance the precision and flexibility of management decisions, balancing long–term planning with immediate operational needs.

D.

Integration with DSS: For the proposed models to be effectively utilized in real–world applications, integration with existing DSSs used by reservoir management authorities is essential. Future research should focus on implementing and testing this integration in operational environments, ensuring that these models can provide practical support for sustainable water resource management. The process involves several key steps, including data integration, model execution, output interpretation, and decision–making support.

E.

Adaptive Management Approaches: Future research should also explore adaptive management approaches that can incorporate near–future information and adjust strategies accordingly. This is particularly important in the face of uncertainties related to climate change and socio–economic developments. Effective water demand management involves iterative learning processes that can inform and implement flexible, responsive policies to enhance the overall resilience and effectiveness of water resource allocation.