Adaptive Neuro-Fuzzy Optimization of Reservoir Operations Under Climate Variability in the Chao Phraya River Basin

Abstract

Reservoir operations play a pivotal role in shaping the flow regime of the Chao Phraya River Basin (CPRB), where two major reservoirs exert substantial hydrological influence. Despite ongoing efforts to manage water resources effectively, current operational strategies often lack the adaptability required to address the compounded uncertainties of climate change and increasing water demands. This research addresses this critical gap by developing an optimization model for reservoir operation that explicitly incorporates climate variability. An Adaptive Neuro-Fuzzy Inference System (ANFIS) was employed using four fundamental inputs: reservoir inflow, storage, rainfall, and water demands. Daily resolution data from 2000 to 2012 were used, with 2005–2012 selected for training due to the inclusion of multiple extreme hydrological events, including the 2011 flood, which enriched the model’s learning capability. The period 2000–2004 was reserved for testing to independently assess model generalizability. Eight types of membership functions (MFs) were tested to determine the most suitable configuration, with the trapezoidal MF selected for its favorable performance. The optimized models achieved Nash-Sutcliffe efficiency (NSE) values of 0.43 and 0.47, R² values of 0.59 and 0.50, and RMSE values of 77.64 and 89.32 for Bhumibol and Sirikit Dams, respectively. The model enables the evaluation of both dam operations and climate change impacts on downstream discharges. Key findings highlight the importance of adaptive reservoir management by identifying optimal water release timings and corresponding daily release-storage ratios. The proposed approach contributes a novel, data-driven framework that enhances decision-making for integrated water resources management under changing climatic conditions.

1. Introduction

An increase in floods and droughts has been seen in the last two decades due to climate change affecting the rainfall patterns in the Chao Phraya River Basin (CPRB) after 1980 [1]. As experiences from previous floods and droughts show, climate change is mostly a problem [2]. In order to obtain solutions for them, climate change models involving extreme rainfall events should be researched together with consideration and a plan for how to prevent and mitigate both flood and drought events in the future for the CPRB. Large-scale reservoirs play an essential role in effective water management and climate change mitigation. Since mis-operation can create multiple negative impacts, an overview of the water discharge is required, taking into account several variables, e.g., available water in dams, water demand from the downstream, and unsystematic hydrological situations. During different times and hydrological characteristics, the dams were operated differently between dry and wet seasons. In the wet season, high inflow, low water demand, flood protection, and saving water for the next dry season are important factors regulating how dams should release water. On the other hand, distributing water depending on the water demands and using the available water in the dams to mitigate droughts is a challenge during the dry season.

Due to the complicated criteria, it is recommended to use computing techniques such as Fuzzy logic, Adaptive Neuro-Fuzzy Inference System (ANFIS), and Artificial Neural Network (ANN) to predict water discharge in the future [3]. The techniques have their pros and cons. The ANN has difficulties in constructing a suitable training algorithm. Fuzzy logic shows good performance in handling uncertainty, non-linear, and noisy data [4]. However, expert knowledge is required to design fuzzy logic rules in order to construct fuzzy modeling [5]. To overcome this issue, the Neuro-Fuzzy System has become more popular due to the use of the neural network to learn an algorithm and the fuzzy system to match input variables associated with output(s) [6,7].

By using the Adaptive Neuro-Fuzzy Inference System (ANFIS), the possibility to tune the inputs in the ANFIS is a key to improving the model performance. In order to simulate reservoir water discharge and storage, two algorithms have been considered: a simulation-based algorithm and an optimization-based algorithm. The simulation-based algorithm requires observed information. Schemes have been introduced to use storage capacity, inflow, and downstream water demand as input variables [8,9]. Additionally, for the optimization-based algorithm, extreme future inflows, rainfalls, and water demands are required in order to decide optimal water discharge and effective reservoir operations [9,10].

Its application in reservoir inflow forecasting has proven effective, particularly when combined with metaheuristic optimization techniques. For instance, Alquraish et al [11] demonstrated that an ANFIS model enhanced with Genetic Algorithm (ANFIS-GA) achieved superior accuracy in forecasting monthly inflows to the King Fahd Dam, outperforming both Hidden Markov Models and Support Vector Machines in terms of RMSE and Nash–Sutcliffe efficiency. Similarly, Tong et al. [12] developed a Copula-PSO-ANFIS model to predict concrete dam displacement, showing that the integration of Copula theory for input selection and PSO for parameter optimization led to approximately 46% lower prediction errors compared to conventional methods. These studies highlight ANFIS’s strength in integrating expert knowledge with data-driven learning to improve predictive performance.

Because of the benefits of the Neuro-Fuzzy system, we have considered using the ANFIS in this research so that the actual dam operation criteria could be formulated through fuzzy rules. Regarding the abovementioned advantages of Fuzzy logic, however, two difficulties in order to construct fuzzy modeling have been discussed [5]. First, expert knowledge and recorded operation rules are required to transform into fuzzy rules for the premises [13]. Secondly, the way users could determine membership functions (MFs) in terms of their values and shapes is a significant approach to maximizing modeling performance.

However, challenges remain in real-time deployment, model generalization, and integration with broader environmental and socio-economic systems, as noted in recent reviews on machine learning applications in reservoir operation. Thus, while ANFIS shows substantial promise, further research is needed to enhance its scalability and adaptability in operational water resource management. In this research, explanations on how to overcome those questions are provided. Therefore, optimal ANFIS model performances for the Bhumibol and Sirikit Dams are obtained by three research aims: (1) to maximize the ANFIS model performance by determining the appropriate form and the numbers of membership functions (MFs), (2) to explore the relationship between the complexity of input variables (inflow, storage, rainfall, water demand, and additional set-up) and output (reservoir releases), (3) to predict reservoir water releases for the future, (4) and to predict the impacts of dam operation and climate change.

2. Methodology

2.1. Study Area

The Chao Phraya River Basin (CPRB) covers an area of 156,618 km², accounting for 31% of Thailand’s area. Their geographical coordinates are from 98°3′53.92″ to 101°31′20.51″ East longitude and from 13°32′22.44″ to 19°48′38.12″ Northern latitude (Figure 1). The CPRB includes five major tributaries—Ping, Wang, Yom, Nan, and Pasak rivers. The Tha Chin River splits off the mainstream at about 80 km downstream of the CPR origin.

The hydrological balance of the CPRB is largely driven by seasonal variations, with rainfall patterns shaped by the monsoonal climate. The rainy season extends from mid-May to mid-October, dominated by moist air masses from the Indian Ocean under the influence of the Southwest Monsoon. In contrast, the dry season spans from mid-October to mid-May and is characterized by cool, dry air brought by the Northeast Monsoon. Annual precipitation across the basin varies from approximately 1000 mm in the lower plains to 1400 mm in the upper mountainous regions, contributing to high spatial variability in runoff and streamflow.

Hydrologically, the tributaries exhibit distinct outflow regimes. The Ping River, originating in mountainous northern Thailand, contributes to early-season flows driven by orographic rainfall and plays a critical role in filling Bhumibol Dam. The Yom River, characterized by steep terrain and the absence of major storage infrastructure, is known for its flashy and unregulated flows, making it a significant contributor to downstream flood risks. The Nan River, regulated by Sirikit Dam, has a more stable and attenuated hydrograph, particularly in the dry season. The Wang River presents more moderate flow dynamics due to its gentler topography. The Pasak River, also regulated by the Pasak Jolasid Dam, contributes flow regulation functions to the lower basin. These tributaries converge in Nakhon Sawan, marking the official beginning of the Chao Phraya River, where flow regimes are affected by upstream dam operations, local rainfall, and tributary interactions.

Significantly, two large-scale reservoirs play an important role in preventing floods during the rainy season as well as supporting the water for agriculture during the dry season in which Bhumibol Dam is located in Ping River Basin and Sirikit Dam in Nan River Basin as shown in Figure 1.

2.2. Reservoir Operations

Bhumibol Dam and Sirikit Dam are being operated by the Royal Irrigation Department (RID) and the Electricity Generating Authority of Thailand (EGAT) together. The origin of the Chao Phraya River is located approximately 240 km downstream from Bhumibol Dam and 360 km downstream from Sirikit Dam in Figure 1. The numerous significant purposes of the two dams include flood and drought mitigation, water supply for downstream, hydro-electrical energy generation, salt intrusion control, and maintenance of the environmental flow during the dry season. The general description of the dams is shown in

Table 1. General description of Bhumibol and Sirikit Dams (from EGAT and RID).

Description	Unit	Bhumibol Dam	Sirikit Dam
Latitude	-	17°14′31″ N	17°46′05″ N
Longitude	-	98°58′31″ E	100°33′15″ E
Drainage Area	km2	26,386	13,130
Storage at Maximum High-water Level	million cubic meters (MCM)	13,462	10,508
Maximum High-water Level	m	260	166
Normal High-water Level	m	260	162
Minimum Water Level	m	213	128
Average Annual Inflow	million cubic meters (MCM)	5704 (1965–2010)	5638 (1974–2010)
Date of Completion of Construction Works	-	June 1964	July 1974

[14].

In general, the dam operation is to serve and store water during the rainy season, especially from August to October. Then, the dams release the water for agricultural demands during the dry season. However, the dams have some criteria to control the water volume between the Upper Rule Curve (URC) and Lower Rule Curve (LRC). In general, the rule curves are adjusted to balance the demand for agriculture, consumption, ecological conservation, and flood protection. In the CPRB, the dam operation plays an essential role in flood control. Previous rule curve rules are inappropriate because it is the one factor that caused the devastating impacts of the floods in 2011. Accordingly, in 2012, new rule curves for both Bhumibol and Sirikit Dams were designed to accommodate a larger water storage volume during the rainy season. Notably, the new URC has been adjusted to a lower value than the old rule curves in the middle of the rainy season. Given the purposes of flood protection and mitigation, dam operations are consequently a key approach to water resources management.

2.3. Data

In order to construct the ANFIS model of reservoir operation, four input variables (inflow, rainfall, storage, and water demand) and an output (reservoir discharge) are set to simulate the dam water discharge at a daily time interval.

The daily reservoir inflow, rainfall, storage, and reservoir discharge of Bhumibol Dam and Sirikit Dam are gained from observed data of the years 2000–2012 collected by EGAT and the Thai Meteorological Department (TMD), respectively. The daily water demands in the boundaries of their irrigation areas are calculated by the Water Requirements Model which includes crop coefficient, moisture status, effective rainfall, and so on, based on local conditions.

For the preparation of future climate data, the Database for Policy Decision-Making for Future Climate Change (d4PDF) using the +4K future climate simulation was prepared for six ensembles (CC, GF, HA, MI, MP, and MR) of two periods: 2051–2070 and 2091–2110. The d4PDF model was applied using outputs from a global atmospheric model with a horizontal grid spacing of 60 km, based on the Atmospheric General Circulation Model 3.2 (AGCM3.2). Simulations were conducted by setting the lower boundary conditions using observed monthly mean sea surface temperature (SST), sea ice concentration (SIC), and climatological monthly sea ice thickness (SIT). Six SST warming patterns were applied using different scaling factors.

The d4PDF rainfall data were corrected by a post-processing technique (interpolation and developed bias correction methods) as shown in Figure 2 in the previous study [14]. The daily reservoir inflow at the upper parts of Bhumibol Dam and Sirikit Dam has been simulated by using the Hydrological Simulation Program-Fortran (HSPF). The storage is calculated by the water balance equation, as referred to in Equation (1), in which the evaporation was calculated by the maximum and minimum temperatures of the d4PDF data by the Hamon method [15].

2.4. ANFIS Model Developments

2.4.1. ANFIS Model Construction

The Neuro-Fuzzy System is designed to resolve the difficulty of “IF-THEN” rules’ construction by taking advantage of fuzzy logic and neural networks in a single structure. Consequently, the Neuro-Fuzzy System uses the neural network to learn algorithms and uses the fuzzy system to capture input variables linked with output(s) [16]. Four processes (fuzzification, firing strength, consequences, and aggregation) contribute to generating the output(s) [9,17]. Fuzzification is a step to identify membership functions (MFs) that ANFIS could determine membership degrees (MDs) in each input variable, as shown in Figure 3. In the firing strength step, the combination of membership degrees could calculate the weight of each rule based on its linguistic label [18]. The consequence is estimated in each rule through the relationship between input variables and output using the least-squares estimator, which can be adjusted during the training process. The overall outputs from the fourth layer are combined as a final output in the aggregation step.

Due to the research objective to maximize the model performances for both dams, membership functions (MFs) were considered in the accomplishment of ANFIS models [6]. Thus, the forms and numbers of MFs were adjusted for four stages as shown in Figure 4.

In the first step of the model construction, due to computer running time, four inputs consisting of inflow, storage, rainfall, and water demand, and adjusting the MFs numbers between n = 3 and n = 5 for each input were set. Since the general form of MFs has a high impact on the model performances, however, no evidence and guidelines are provided on which form of MFs should be used to set up ANFIS models for dam operations and water resources. In previous studies, the Generalized bell (gbellmf) shaped membership functions have been used without testing [9]. There was the question: which form is most appropriate for the use of dam operations? To investigate a suitable form of MFs, we adjusted eight MF forms: Generalized bell (gbellmf), Gaussian (gaussmf), Gaussian combination (gauss2mf), difference between two sigmoidal (dsigmf), product of two sigmoidal (psigmf), Pi (pimf), triangular (trimf), and trapezoidal (trapmf) [16].

After selecting the proper form for the MFs, in the second stage, we have set three inputs (inflow, storage, rainfall) and four inputs, adding water demand. The MFs numbers for inflow and storage varied from n = 3 to n = 8, while the numbers between n = 3 and n = 5 are for rainfall and water demand.

Coerver et al. [9] has mentioned that seasonality improves model performance. In the third stage, the four input variables used in the previous setting were separated into two datasets: the rainy season dataset (May–October) and the dry season dataset (November–April). Then, the seasonal ANFIS models were independently constructed using both datasets and adjusting various numbers of MFs following the previous range of MF numbers. The suitable seasonal ANFIS models for dry and rainy seasons were combined to simulate the daily reservoir discharge.

In order to consider monthly dam operation, the fourth stage is to add an input variable indicating the month. Consequently, five input variables were operated with the selected MFs form in the first stage with adjustable numbers of MFs: from n = 3 to n = 8 for inflow and storage inputs and from n = 3 to n = 5 for rainfall, water demand, and month inputs. However, one more input variable increases the time consumed by model construction, especially when the model has more complex membership functions (MFs). Due to this issue, the MFs’ numbers for the monthly input were compared by their performance, while the MFs’ numbers for the other inputs were fixed. At a chosen number of MFs for the monthly input, the numbers of the MFs for the four input variables were randomly changed to construct ANFIS models.

2.4.2. Dam Operation Policy

Due to the limitations of capacity, maximum water release of a reservoir, and downstream river capacity, the dam operation policy was recommended to be carried out with the simulations of the ANFIS models [3]. In the models, the current water release and reservoir storage were inputs to calculate the storage at the next time interval by a general reservoir water balance equation (Equation (1)).

$$S_{t+1}=S_t+Q_t+R_t-O_t-E_t$$

(1)

The storage at the next time interval $\left(S_{t+1}\right)$ is calculated by the inflow ($Q_t$), the rainfall ($R_t$), the evaporation ($E_t$), the storage at the current time interval ($S_t)$ and the released water ($O_t$) [19]. In Equation (1), the released water ($O_t$) estimated by the ANFIS model states that the amount of water, in general, depends on the settings of the membership functions (MFs). However, the simulated water release can be outside of the range between maximum and minimum water releases because of noise. Therefore, the water releases of Bhumibol and Sirikit Dams were commanded to follow three operation rules (Equations (2)–(4)) during the simulation as follows:

The water release ($O_t)$ should not be less than 0 m³/s. If the simulated water release is less than 0 m³/s, the release is set to be 50 m³/s as the minimum reservoir water release for Bhumibol and Sirikit Dams in Equation (2).

$$O_t<0{\mathrm{m}}^3/\mathrm{s},\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n}{O}_t=50{\mathrm{m}}^3/\mathrm{s}$$

(2)

If the storage at the next time interval is greater than the maximum reservoir’s capacity ${(S}_{max})$, then the water release at the current time interval ($O_t$) is re-calculated:

$$S_t>S_{max},\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n}{O}_t=S_t{+Q_t+E}_t-S_{max}-R_t$$

(3)

If the simulated water release ($O_t$) is greater than 1200 m³/s, we set-up the water release equal to 750 m³/s and 735 m³/s for the Bhumibol and Sirikit Dams, respectively.

$$O_t>1200{\mathrm{m}}^3/\mathrm{s},\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n}{O}_t=750{\mathrm{m}}^3/\mathrm{s}\mathrm{a}\mathrm{n}\mathrm{d}735{\mathrm{m}}^3/\mathrm{s}$$

(4)

This was gained from their maximum water release rate from their construction until 2016.

2.4.3. Model Performance

In combination with dam operation policy and various model setups, the reservoir water releases were simulated for both Bhumibol and Sirikit Dams. Since noise in the outputs cannot be represented during the training process, a testing phase is necessary to detect potential overfitting.

To evaluate the performance of the ANFIS model in simulating reservoir water releases for the Bhumibol and Sirikit Dams, the dataset was divided into two distinct periods. Although it is more common to use earlier years for training and later years for testing in time series modeling, we adopted the reverse approach in this study for specific reasons. The period from 2005 to 2012, which includes several extreme flood and drought events, was selected for training to ensure the model is exposed to a wide range of hydrological conditions and can learn more complex operational patterns. The earlier period, 2000–2004, was then used for testing to independently evaluate the model’s generalization capability under relatively moderate conditions. This selection was made to prioritize model robustness and to reflect operational scenarios where a model trained on extreme events is assessed against more typical conditions, while also maintaining the temporal integrity of the dataset. Finally, the performance of the simulations in each model with a different set-up is identified by Nash Sutcliffe efficiency (NSE), the coefficient of determination (R²), and the root mean square error (RMSE).

$$NSE=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(O_i-P_i\right)}^2}{{\sum }_{i=1}^n{\left(O_i-\overline{O}\right)}^2}}$$

(5)

$$R^2={\displaystyle \frac{{\sum }_{i=1}^n(O_i-\overline{O})(P_i-\overline{P})}{\sqrt{{\sum }_{i=1}^n{\left(O_i-\overline{O}\right)}^2}·\sqrt{{\sum }_{i=1}^n{\left(P_i-\overline{P}\right)}^2}}}$$

(6)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(O_i-P_i\right)}^2}$$

(7)

where $O_i$ is the observed reservoir discharge at time step i, $P_i$ is the predicted reservoir discharge at time step i, $\overline{O}$ is the mean of observed discharge values, $\overline{P}$ is the mean of predicted discharge values, and $n$ is the total number of observations.

2.4.4. Additional Reservoir Operation for Flood and Drought Protections

Since the ANFIS model only learns the relationship between inputs and outputs from historical data, the model does not include special dam operations, such as protection against future floods and droughts. In order to achieve an effective reservoir operation, the rainfall downstream of Bhumibol and Sirikit Dams is taken into account in how the dams should release water. The daily rainfall data at ten rain gauge stations, in which three, five, and two gauge stations are located downstream of Bhumibol Dam, Sirikit Dam, and Lower CPRB, respectively, were estimated by the 90th percentile value of the rainfall of each station. The operation rules for flood prevention were designed by considering the timing and the daily release storage ratio ($r$). The daily release storage ratio ($r$) is calculated by the reservoir release divided by the storage at the current time interval. The ratios have been examined for the appropriate range, which is not explained in detail in this paper. For the additional operation, the ratios ($r$) of 0.002 and 0.006 were adjusted to be additional water releases between July and August under the conditions: the downstream rainfall is smaller than their 90th percentile values, the storage ($S_t$) is greater than 8000 million cubic meters (MCM) for the Bhumibol Dam and 6000 MCM for the Sirikit Dam, and the simulated water release via the ANFIS model ($O_{t,ANFIS}$) is lower than 52 MCM as calculated by Equation (8).

$$O_t=O_{t,ANFIS}+r\times S_t\hspace{1em}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{J}\mathrm{u}\mathrm{l}\mathrm{y}–\mathrm{A}\mathrm{u}\mathrm{g}\mathrm{u}\mathrm{s}\mathrm{t}$$

(8)

The daily release storage ratio ($r$) of 0.001 was commanded for the reservoir release to mitigate downstream overflow during September and October when the simulated water release via the ANFIS model ($O_{t,ANFIS}$) is higher than 4.3 MCM, and the rainfall is greater than 90th percentile values as in Equation (9).

$$O_t=0.001\times S_t\hspace{1em}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{S}\mathrm{e}\mathrm{p}\mathrm{t}\mathrm{e}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}–\mathrm{O}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{b}\mathrm{e}\mathrm{r}$$

(9)

Reservoir operation rule for drought mitigation was considered after more than 7 consecutive dry days during dry seasons (November–April) in the Lower CPRB. Because the lowest water volume of 7 MCM is necessary for the Chao Phraya River, Bhumibol and Sirikit Dams were set to release additional water volumes of 14 MCM and 7 MCM, respectively, as in Equations (10) and (11).

$$O_t=O_{t,ANFIS}+14\mathrm{M}\mathrm{C}\mathrm{M}\hspace{1em}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{B}\mathrm{h}\mathrm{u}\mathrm{m}\mathrm{i}\mathrm{b}\mathrm{o}\mathrm{l}\mathrm{D}\mathrm{a}\mathrm{m}$$

(10)

$$O_t=O_{t,ANFIS}+7\mathrm{MCM}\hspace{1em}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{S}\mathrm{i}\mathrm{r}\mathrm{i}\mathrm{k}\mathrm{i}\mathrm{t}\mathrm{D}\mathrm{a}\mathrm{m}$$

(11)

In order to choose an appropriate daily release storage ratio ($r$) for the reservoir operation during July and August, the impacts of the operation using the ratios are examined on river flows. The simulated daily reservoir releases through the ANFIS models and the additional operation were inputs in the hydrological model, as explained in Section 3.1. Daily river flows at the Chao Phraya River origin are simulated for three periods: 2000–2010, 2051–2070, and 2091–2100. Then, the changes in river discharges are also represented by the non-exceedance probability of the annual maximum daily flow.

After the ANFIS model performance and the appropriate daily release storage ratio ($r$) were confirmed, the future data were generated by the ANFIS model, and the additional reservoir operation for future daily reservoir releases of the three ensembles for the years 2051–2070 and 2091–2110.

3. Results and Discussion

The three objectives of this paper, the selection of a suitable form, and the numbers of the membership functions (MFs) are discussed in the section of ANFIS model development for both Bhumibol and Sirikit Dams. Also, the additional dam operation and the trends of water releases in the future are investigated and suggested.

3.1. ANFIS Model Development

3.1.1. Selecting a Form of Membership Functions (MFs)

Several studies have reported that the required training and testing processes are time-consuming due to trials and errors [7,9,20]. In order to save time, the generalized bell-shaped membership function is generally used for reservoir operation without the comparison of other forms. A few studies have recommended which form of a membership function should be used to generate daily reservoir releases. In order to optimize the model performance, eight forms of membership functions were used during the training process using a similarly observed dataset of the years 2006–2012. The numbers of MFs associated with each input variable were adjusted to n = 3 based on linguistic values that are high, medium, and low.

With regard to the comparison of different forms, Figure 5 illustrates model performances via the coefficient of determination (R²) for Bhumibol and Sirikit Dams. The two sigmoidal types (dsigmf and psigmf), Generalized bell (gbellmf), Gaussian (gaussmf), and Gaussian combination (gauss2mf) express low correlation between observed data and simulated results, which means R² values are near zero. The triangular-shaped (trimf) and the pi-shaped (pimf) membership functions could increase the model performance; however, R² values are mostly lower than 0.5. The trapezoidal-shaped form (trapmf) indicates an acceptable simulation quality (R² > 0.5) on specific numbers of MFs for both dams. Considering the number of MFs used for the trapmf form as shown in Figure 6, the R² values obviously increase when the numbers of MFs for inflow and storage as input variables are more complex [21].

3.1.2. Impacts of Input Variables

Since using the trapezoidal-shaped form could improve the model performance, more input variables are added for both Bhumibol and Sirikit Dams in order to consider the multiple factors such as hydrological variables, human activities, and timing. However, by increasing one input variable, the model construction of each set of membership functions requires more time because its fuzzy system needs to add more fuzzy rules and tune more premise and consequence parameters. Accordingly, this section aims to determine which input variables are significant to improve model performance.

The maximum NSE values of ANFIS models during the training process are 0.75 for the Bhumibol Dam and 0.86 for the Sirikit Dam. During the checking process (Figure 7b), adding only the WD input could slightly reduce model performance. Combining the inputs of WD with the month could produce the highest NSE values of 0.42 and 0.48 for Bhumibol and Sirikit Dams, respectively. The better performances via the input indicating the month are caused by two reasons: (1) the dam operations are generally based on the monthly pattern, and (2) additional inputs increase fuzzy rules. For example, the basic parameters have created 270 fuzzy rules, while the combinations between basic and additional inputs have established 540 rules. Thus, the increase in fuzzy rules could have additional potential to cover rare cases of dam operation, including flood events [22].

Figure 8 shows the results of NSE values by increasing the input variables, in which water demand (WD) and inputs identifying season and month were added from the basic input variables (inflow, storage, and rainfall). Regarding the training process (Figure 8a), NSE values could be increased by adding more input variables; however, the reduced model performances of Bhumibol Dam are expressed by including the inputs for WD and season.

3.1.3. Impacts of Numbers of Membership Functions (MFs)

As reported in [16], the performance of the ANFIS models could also be enhanced by increasing the numbers of MFs. Therefore, in this study, we performed systematic membership by ranging it between n = 3 and n = 8 for inputs of inflow and storage and from n = 3 to n = 5 for rainfall, water demand, and month were adjusted for ANFIS models of Bhumibol and Sirikit Dams, as shown in Figure 8 and Figure 9. Due to the different settings, 2000 models for the Bhumibol Dam and 2000 models for the Sirikit Dam have been constructed. Because more complicated adjustments took a longer time to create the models, n = 3 and n = 5 for the month were not simulated when other parameters were varied.

In the training process (Figure 8), the NSE values could be increased by adding a higher numbers of MFs for inflow and storage inputs. Particularly, when the number of MFs for storage is 8 for Bhumibol Dam (Figure 8a), the NSE values are higher than 0.7. In the latter of Sirikit Dam (Figure 8b), the NSE values are increased by increasing the number of MFs for storage; however, its number of 8 reduces the NSE value. Moreover, the numbers of MFs for inflow are increased, and the NSE values tend to increase slightly for both dams. In contrast, more complex numbers of MFs for rainfall and water demand are likely to reduce the NSE values. Considering the monthly input, a higher number of MFs could better simulate more fit to observed data (blue dots in Figure 8). Consequently, it can be implied that the MF number for the storage parameter is the most significant in improving model performance.

In general, when the training and testing processes were operated, the ANFIS model could select parameters based on the minimum testing data error, which is always included during noisy measurements. Since the noise of data could not be represented during the training process [20], in there, the NSE value of the testing process was prioritized to optimize the model performances of Bhumibol and Sirikit Dams as the results shown in Figure 9 [13].

• Bhumibol Dam

The relationship between the NSE values and input variables could not be concluded for the ANFIS model testing process of Bhumibol Dam, as shown in Figure 10a. The optimized performance of the ANFIS model of Bhumibol Dam is 0.43 of NSE value (R² = 0.59 and RMSE = 77.64) when the MF numbers for inflow, storage, rainfall, water demand, and monthly inputs are n = 6, 8, 3, 3, and 3, respectively. The simulated daily reservoir discharge of the optimized Bhumibol Dam model performance is shown for the training and testing processes in Figure 10. In the training period of the years 2005–2012, the trend of water discharges is similar for observation and simulated output. About 50–400 m³/s of water discharge during the dry season. Particularly, in 2011 and 2012, more than 400 m³/s of water was attempted to be discharged from the Bhumibol Dam, the highest value since the 2011 flood condition (Figure 10a).

However, the ANFIS could not capture this special operation since downstream floods are not considered in the model; therefore, the simulated result is underestimated in the training process. In the testing process from the years 2000 to 2004 (Figure 10b), the daily water releases are generally overestimated. This result could imply that during the training process, the premise and consequence parameters are determined by learning the dam operation pattern. However, the dam operation of Bhumibol Dam is unrestricted. Consequently, the change in MFs numbers and the additional parameters by setting water demand and month could improve the model performance, but not enough to obtain good results for the daily time series.

b.

Sirikit Dam

Since the more complicated numbers of MFs for inflow, storage, rainfall, and water demand tend to decrease the model performance of Sirikit Dam as shown in Figure 7b; accordingly, the maximum NSE value is 0.47 (R² = 0.50 and RMSE = 89.32) when the MFs numbers for inflow, storage, rainfall, water demand and month inputs are n = 4, 3, 3, 3, and 4, respectively. The reduced complexity of the MFs number means that the Sirikit Dam was operated with similar operation patterns year after year. Particularly, the MFs number for storage input at n = 3 could be implied in the linguistic label into low, medium, and high storage.

The higher complexity of the inputs and the member functions seems to be insufficient in order to obtain good performance for simulating the daily reservoir releases. It is caused by the limitation of the timeframe for the training period. The duration should be extended to include more cases of floods and droughts [9]. Additionally, the daily water releases of the training and testing processes are presented for the dry and rainy seasonal datasets, as shown in Figure 11.

The optimized ANFIS model with five parameters could learn an algorithm for Sirikit’s Dam operation since the dam has a similar pattern of water discharge almost every year. Exceptions are the high water discharges of the dry seasons in 2011 and 2012 due to the high water storage after the massive floods in 2011. During the testing process, the optimized ANFIS model could simulate daily water releases, especially during the dry season of the normal water years 2002–2004. The simulated water discharges during the rainy season of the year 2001 fluctuate due to abnormal dam operations, such as the sudden release of high water amounts over 500 m³/s.

3.2. Additional Dam Operation

Since the ANFIS models of both dams could not capture the special reservoir operation during the floods in 2011, the timing and the daily release storage ratios ($r$) were applied for the additional operation. In general, when the annual maximum river flow occurs in September or October, the dam operation in both months in the model is commanded to release a small amount of 0.001 of the current storage. By releasing the small volume at the end of the rainy season, the reservoirs are required to have bigger storage and release a higher amount of water.

The daily release storage ratios ($r$) of 0.002 and 0.006, accordingly, were adjusted to increase the water discharge during the mid-rainy season as shown in Figure 12. Moreover, the additional operation managed to add an additional water release of about 14 MCM for the Bhumibol Dam and 7 MCM for the Sirikit Dam during the dry season. Through the set-up of the daily release storage ratios, the reservoir water release is slightly higher between July and August, and very high water releases (>600 m³/s) are dramatically decreased (Figure 12a). Also, an increase in the water release is shown during the dry season (November–March).

In order to choose the appropriate daily release storage ratio, the impacts of both ratios (0.002 and 0.006) were examined on the daily river flow at the origin of the Chao Phraya River. The river flow of the period 2000–2010 (Figure 13a) is slightly increased during the low flow. In contrast, the highest flows are decreased, particularly for the flood peaks in 2002 and 2006.

In future maximum flow, there is a trend of a general increase as shown in Figure 13b. This river flow could be damped out by using the daily release storage ratios of 0.002 and 0.006. Particularly, the highest decrease in the flow peaks is identified by the ratio of 0.006. Moreover, the non-exceedance probability of annual maximum daily river flow using the daily release storage ratio (r) of 0.006 also confirmed the decrease in extreme flow events, as shown in Figure 14. When the non-exceedance probability is higher than 0.50, the annual highest flow could be reduced by about 500–1500 m³/s from the ANFIS model. Consequently, the daily release storage ratio of 0.006 is the best choice to simulate future reservoir water release for flood prevention.

4. Conclusions

This research aims to construct ANFIS models to simulate daily reservoir releases for Bhumibol and Sirikit Dams. In order to optimize the performance of the model, input parameters, forms, and numbers of MFs, and additional set-ups have been taken into account. A suitable form of membership functions (MFs) has been selected as the first priority. From the eight forms of MFs, it was found that the trapezoidal-shaped membership function leads to the best result for constructing a reservoir operation model. Then, basic input parameters (inflow, storage, rainfall, and water demand) and additional considerations of seasonality and the month have been applied.

For the basic parameters, the adjustment of the numbers of MFs for inflow and storage inputs has a significant improvement in model performance. In contrast, more complexity of numbers of MFs for rainfall and water demand inputs does not tend to improve the performance. The set-up of seasonality could increase the performance, especially for the ANFIS model of Sirikit Dam, because the dam operation is based on the seasonal pattern. With regard to additional monthly input, the model performance of Bhumibol Dam is greatly increased since the water discharge depends on monthly antecedent effective storage ratios. Consequently, an additional monthly parameter expressed the best performances, leading to NSE values of 0.43 and 0.47 for Bhumibol Dams and Sirikit Dam, respectively.

In order to achieve an effective reservoir operation, additional reservoir operations have been suggested for the timing and the daily release storage ratios ($r$). The simulated water release by adding the daily release storage ratio ($r$) of 0.006 is recommended for use at the mid-rainy season (July–August), and the ratio ($r$) of 0.001 should be applied at the end of the rainy season (September–October). These ratios were examined for their impacts on river flow, and the results show that the flood peaks could be decreased. In consideration of drought mitigation, the operation was designed to increase the daily water release of 14 MCM for the Bhumibol Dam and 7 MCM for the Sirikit Dam when the Lower CPRB had no rainfall for over seven days.

The optimized performances of ANFIS models and the additional reservoir operation for both dams have been applied to future climate data of two periods (2051–2070 and 2091–2110). The dams discharge about 700–1200 MCM per month during the dry season. About 500–2000 MCM per month and less than 400 MCM per month are released at the middle and the end of the rainy seasons, respectively. In total, through the rainy season, about 6000–8000 MCM could be stored to support water downstream. In the future, the available water from both dams should be taken into account to meet water demands for effective dam operations.

The unavailability of consistent post-2012 data for key input variables across both dam sites restricted the model validation under future hydrological conditions. As a result, the long-term predictive performance of the ANFIS model under evolving climate and demand scenarios could not be fully assessed. Future research should focus on incorporating more recent datasets, including projected climate and demand trends, to evaluate the model’s robustness in forecasting reservoir operations under changing environmental conditions. Additionally, integrating ensemble forecasting techniques or coupling with climate models could enhance the model’s adaptability and reliability in long-term water resource planning.