Assessment of sustainability index for rural water management using ANN Open Access

The word sustainability links to various meanings in different contexts. One way to define sustainability is to consider economic, social, and environmental factors for the overall development and impact of a tool or a product (Sala et al. 2015). Contemporary innovations focus on sustainability issues in engineering applications. Since the development of tools or products needs to follow quantified parameters from a sustainability perspective, various measures such as the indicators, audits, standards, benchmarks, certificate systems, appraisal, and sustainability indexes have consequently come into effect (Dalal-Clayton & Sadler 2014). Simultaneously, under the Sustainability Development Goals – 2030 set by the United Nations, many developing countries have been employing sustainability index models to achieve sustainable society status (Van Beynen et al. 2018). Hence, this study focuses on the sustainability aspect by incorporating engineering concepts in rural water management.

Current research intends to assess a water management system through a sustainable approach. Previous studies in the extant literature have already applied several techniques in evaluating water management systems. Some of these techniques are multi-criteria analysis, external costs, energy analysis, ecological footprint methods, and energy life cycle analysis (Willis & Garrod 1998; Lundie et al. 2004; Cabrera et al. 2010; Herstein et al. 2011; Shrestha et al. 2012). For instance, Willis & Garrod (1998) conducted a study on the external costs of water extraction from rivers and aquifers to provide a perspective on maximising society's benefits. Cabrera et al. (2010) assessed water networks in an urban scenario with the scope to compare and improve different networks' energy efficiency. Shrestha et al. (2012) reported that water distribution from valleys or lakes has carbon emission during water conveyance to different locations. As a result, an assessment of water distribution is crucial in the reduction of CO₂ emissions. The application of heuristic methods evaluating water distribution systems is gathering attention in the current times. Also, optimisation using machine learning systems on the water management system perceives a great understanding of life cycle analysis (Herstein et al. 2011). The literature review for this paper on various assessment techniques on different water management systems helped identify a sustainable approach.

Among all the above techniques, Life Cycle Analysis (LCA) is one of the most popular practices employed in making methodological decisions for developing sustainability strategies. Application of LCA on estimating a water system's operating performance acts as a communication device to assess alternative future scenarios for strategic planning. Lundie et al. (2004) developed an LCA model on Sydney Water's total operations for 2021 by extrapolating Life Cycle Inventory data from annual performance data. This model helped to determine population growth, demand management, water quality, and treatment assessments well in advance. Over the years, there is progress in advanced strategies for evaluating water resources. Another technique named the Set Pair Analysis method, a fuzzy-based approach, determines the coordination ability of sustainable water resource utilisation and their carrying capacity in regional groundwater (Wang et al. 2009). Simultaneously, participatory terms was a primary concern of rural community water management. In order to overcome this issue, the application of multi-criteria analysis concerning the evaluation of sustainability in rural water supply systems considered semi-structured interviews, social mapping, technical and household surveys, and water monitoring. For example, Domínguez et al. (2019) applied the Multi-Attribute Theory using the Analytic Hierarchy Process (AHP) on a rural water supply system for assessing multiple communities' direct influence and sustainability. The conclusive remarks of their study highlighted the gravity of incorporating multiple attributes into a single aspect besides evaluating the sustainability of the rural water supply system.

The literature study perceives that the focus on sustainable factors through water resources got its prominence due to increased water demand and depletion of the groundwater table. All the water resources planned for the water system's replenishment must aim for a sustainable factor check. Different measures in sustainability have a unique approach. One of them is the sustainability index (SI), meant to simplify the complex decision-making process for setting up a sustainable system. The idea behind the SI aggregates the relative information about the sustainability aspects of various decisions, strategies, and approaches, providing an easy-to-understand scoring system to readily determine the most sustainable options (East 2014).

SI can be determined for different water management scenarios using different performance criteria. Generally, definition of sustainability concerning water resources is proposed as:

It infers that creating a sustainable water resource must target the complete need of society by simultaneously maintaining a balance among the combined effect of ecology, environment and hydrology. McMahon et al. (2006) focused on determining SI in water resources by reservoirs' performance integrated with hypothetical storages located on the rivers and storage-yield-performance analyses. Recent studies also suggest the measurement of performance evaluation by Evidential Reasoning based SI for alternate water management scenarios. For instance, Jahanshahi & Kerachian (2019) determined SI of a whole community, which involved the principle to obtain the Group Sustainability Index (GSI) for different smaller groups. Further, a Brazilian study on water resources planning and management considered SI as a performance evaluative for alternative policies from the perspective of water users, where SI identifies the policies that evaluate, compare, and identify the required water management characteristics (Sandoval-Solis et al. 2011). Also, Cai et al. (2002) evaluated a study, highlighting the development of a long-term revamped model including quantified sustainable criteria for steady water supply to meet crop water demands, municipal and industrial water demands, and environmental conservation. Hence, our review of the extant literature infers those real-life models and solutions behave as guidelines to the decision-making process using sustainability indicators.

Mere implementation of techniques for attaining sustainability indices is neither advisable nor non-beneficial. A comprehensive SI application is lacking in India because of unreliable and unavailable surveyed data, besides location variant socio-economic conditions. A case study on the Indian scenario suggests an expert-based Water Sustainability Index for quantifying water resource sustainability while considering the terms such as natural resources, infrastructure and technology, economic factors, human health, environmental health, and institutional efficiency (Kansal & Gaur 2011). Hence, analysing SI for any condition is dependent on constraints such as specific location and method adopted. The present study attempts to develop an SI metric considering a comprehensive set of factors (Figure 3) affecting a system. Subsequently, the system has been modelled using Artificial Neural Networks (ANN) to account for the system's non-uniformity in different conditions.

ANN is a technique through which decision-makers can present predictions based on the existing historical/observed data. ANN acts as a valuable information processing system that resembles characteristics with a biological neural network of the human brain illustrated by many highly interconnected processing elements called nodes/neurons. These neurons operate in a parallel configuration in ordinary architectures. Each neuron is joined with another by a connection link, where each connection link is related to weights containing information about the input. The neuron net uses the obtained data to solve a particular problem (Sivanandam & Deepa 2014). Hidden layers act as the connection between input and output layers. An interconnection formed by involving multiple layers of computational units in a single direction is known as a multilayer feed-forward neural network (Figure 1). These hidden layers can be partially or fully connected, and training algorithms adjust the associated weights to dismiss the forecasting/prediction error of ANN (Amin-Naseri et al. 2013).

ANN is accepted and applied extensively in multiple research domains worldwide. Application of ANN in the municipal water network of Canadian municipalities benefitted in the performance assessment and prediction of the prevailing water networks' rehabilitation (Al-Barqawi & Zayed 2008). Another literature study infers that an ANN model assisted to analyse and improve the water quality standards in the West Coast of Malaysia (Sulaiman et al. 2019). Hence, previous studies suggest SIs determination concerning the concurrent conditions, whereas this proposed research generates SI for future water management scenarios. The methodology applied for the determination of SI for the future water management scenarios comprises of an integrated approach which includes: (1) the development of a modified SI expression based on water consumption activities and conservation methods; (2) development of an ANN model to determine the SI for new data; and (3) forecasting of the water consumption variables for upcoming years. A flow chart presenting the integrated approach in the study is shown in Figure 2, where the water resource management in a rural area has incorporated a comprehensive framework comprising other factors/indicators like technical, operational, institutional, risk, and climate, apart from economic, environmental, and social factors (Figure 3). Further, the ANN model has been developed to determine the SI by applying dummy variables for water conservation measures, i.e. sprinkler irrigation, drip irrigation, rainfall-runoff storage, where 1 denotes their application and 0 denotes their discard.

Consequently, an ANN model was trained based on the data, with the variables of concern as inputs to the model and their respective sustainability indices as the required target. Since evaluating water consumption and conservation involved many calculations, the model was trained for the readily available inputs, i.e. first SI was derived for the calculated values of water conservation measures. After that, the ANN model was trained for the substituted values of 1's or 0's for the water conservation measures if applied or not applied, respectively. It would help the decision-makers select the water conservation measures without evaluating the watershed area, crop and population water requirement, and conserved water volume. After training the ANN model, it was simulated for a few of the considered future scenarios to derive the study area's respective sustainability indices in the future. One of the simulated future scenarios was selected as an appropriate solution for implementation over the study area to obtain sustainable water resource management.

Data from the columns: population and cultivated area of Table 1 from 2007 to 2014 (eight consecutive years) aided in deriving the respective total annual water consumption over the village. As per the Indian Standard code, the average water per capita is 0.04 m³ for communities up to a population of 20,000 (NBC 2016). The data collected from the survey regarding the required irrigation water quantity (per hectare) for different crops determined crop water requirements. Application of the AHP method in the study area derived the weights associated with the various factors considered. Each factor's weight listed in Table 2, along with evaporation, was derived based on its average value over the eight years, proportioned to the respective average value of water scarcity. The difference between annual water consumption and annual groundwater recharge evaluated annual water scarcity. Since the weight corresponding to the rainwater harvesting system was 0 up to three significant figures after the decimal, it was ignored. The factors and weights obtained (in percentage) are listed in Table 2.

These 40 datasets were used to train the ANN model in MATLAB (hidden layer: 1; output layer: 1; hidden neurons: 10; output neurons: 1; hidden layer transfer function: tansig (Tan-sigmoid); output layer transfer function: purlin; training: Levenberg–Marquardt) (Figure 4). Population, rainfall, evaporation, and cultivation area collected from different sources made the input data readily available. Still, quantification of water conservation measures needs the evaluation of water consumption and rainfall-runoff. Therefore, data corresponding to water conservation measures was substituted by dummy variables, i.e. in case of applying any water conservation method, its input value was taken as 1 else 0 in its absence, as presented in Table 3. Therefore, the trained ANN model was concerned with seven input variables (population, rainfall, evaporation, cultivation area, sprinkler irrigation, drip irrigation, rainfall-runoff storage) and one output variable, SI.

The study considered seven water conservation scenarios as shown in Table 4 with respect to the forecasted values of the above variables. Scenarios S1 and S7 can be viewed as the worst and best-case scenarios since S1 considers minimum and maximum historic rainfall and evaporation, respectively, to continue for future years and vice versa for scenario S7. The rest of the scenarios (S2–S6) were water conservation scenarios as depicted by columns S, D, and R in Table 4. Variables of population and cultivated area were forecasted for all the scenarios (S1–S7) by Equation (4), hence they are not represented in Table 4.

The trained ANN model over the generated data for scenarios T1–T5 was found fit, where it utilised 50, 25, and 25% of datasets for training (R=0.99298), validation (R=0.99875), and testing (R=0.99857), respectively. The overall fitness achieved was R=0.9951(Figure 5(d)). The ANN process in MATLAB normalised the data in the range 0–1 and produced final outputs in the original range. Since the ANN model had a random data set for training, it was repeatedly trained until a mean average error (MAE) below 1% between the evaluated SIs using Equation (3) and the trained ANN model has been achieved.

Variables affecting the water consumption, i.e. population, evaporation, rainfall, and cultivated area from 2008 to 2014, were used for forecasting their respective values from 2015 to 2025 as per Equation (4) (Table 5). The trained ANN model aided in evaluating SIs for scenarios S1–S7 from 2015 to 2025 (Table 6) by using forecasted water consumption activities along with the water conservation measures as per Table 4.

SIs determined for different future scenarios established that water resources were the least sustainable in scenario S1 and most sustainable in scenario S7. Since climatic conditions assumed for scenario S7 were incredibly uncertain, interventions for scenario S6 were planned for full-fledged sustainability. The error between the SIs obtained using Equation (3) and the developed ANN model was evaluated as MAE for the training period 2007–2014. The MAE acquired for the 40 datasets generated over the training period was 0.6%. Loucks et al. (2005) described that sensitivity analysis determines the uncertainty contribution of the variables affecting a system, representing the level of model outputs affected by model inputs. Although ANN modelling is free from any scientific approach specific to the problem, identifying the uncertainty contribution of input parameters for the concerned logical/mathematical model defining the problem system must be considered.

It was evident from the study that scenario S6 was highly suitable as per the ANN modelling of the water resource management problem in the village Kidwana, yet the AHP synthesis model can test the reliability of the results obtained. Again scenario S6 was the best decision to be implemented in the study area, as shown in Table 8. All the factors' values in different scenarios were averaged from 2015 to 2025, which were further normalised with respect to their maximum average value among all the scenarios before placing them into the synthesis model. The second row of Table 8 constitutes the weights of factors obtained as per Table 2. The evaluated effective score was the sum of the products of normalised values of the factors with their respective weights. The higher the effective score, the more suitable the scenario/solution. Though S7 was the most effective, it was incredibly significant due to the maximum rainfall and minimum evaporation for all the future years. Hence, interventions corresponding to S6 would also accommodate S7.

The average catchment area of 25 houses was considered to be 60 m². The average forecasted rainfall (Table 5) from 2015 to 2025 turned out to be 0.4156 m. After assuming the runoff coefficient equal to 0.8, the estimated net runoff was 19.992 m³ per house. Therefore, 25 household tanks corresponding to 25 houses with the potential of storing 20 m³ of water were built (Figure 6). In a circumstance of overflow, the excess water from all the household tanks was routed through a channelled network connected to the community tank, subsequently storing rainwater effectively for future purposes.

This study concludes into two categories covering methodological and practical aspects. The methodological part comprises the process of deriving the sustainability index of a region, while the practical part includes a problem's solution. An ANN model's end-user can easily apply the developed approach for determining SI through this study. The approach avoids experts' involvement in deriving SI after training the ANN model over a specific problem. This method provides an experts' bottom-up approach covering various possible factors affecting a system and stakeholders' top-down approach to analyse the system for different scenarios. The study's practical aspect suggests that the obtained results of SI as per scenario S6 until 2025 attempted to implement drip irrigation and rainfall-runoff storage in the study area at a village. The data considered for the ANN model with seven input variables were study-area dependent. The SI generated could assist as a reference point in decision-making for the rural water management system. The study helped decide the implementation of rainfall-runoff storage and the rainwater harvesting system to provide a sustainable solution. SI derived from the ANN application would help determine sustainability prospects over a region for the present and the future with lesser and readily available information. The household tanks built in the village were appropriately utilised for runoff storage, whereas efforts are being made to construct a community tank to collect the surplus water beyond the capacity of household tanks.

The authors would like to acknowledge the valuable comments and suggestions of the reviewers, which helped to improve this paper's content significantly.

All relevant data are included in the paper or its Supplementary Information.