Search

Water quality index (WQI) has been utilised in many countries and regions as a numeric representation of the condition of water resources. However, the computation of the WQI involves a host of water quality variables. Although machine learning models are proven to be a promising tool to estimate WQI with lesser inputs, sufficient data or samples must be collected so that the machine learning models can be trained well. This exhibits a great challenge in places where there has been a lack of data collection infrastructure to meet the needs of machine learning models. Data scarcity is a major issue to be tackled. This study covered two major rivers that served as water intakes in Peninsular Malaysia (Selangor River and Skudai River), where four synthetic data generation methods, namely the conditional tabular generative adversarial network (CTGAN), the tabular variational autoencoder (TVAE), the Gaussian copula (GC) and the copula generative adversarial network (CopulaGAN), were used to synthesise datasets based on the real dataset. By using the pairwise correlation difference (PCD), Kullback-Leibler divergence (KLD) and the Kolmogorov-Smirnov (KS) test, the best synthetic datasets were selected for the two rivers. The CopulaGAN1 and the CopulaGAN2 yielded the best small and large synthetic datasets at Selangor River, scoring the lowest PCD, KLD and KS statistics. For the Skudai River, the TVAE1 and TVAE2 were chosen. The real and synthetic datasets were used to train the back-propagation neural network (BPNN) for the WQI estimation. Based on the various evaluation metrics, it was proven that increasing the size of training data using the synthetic data method had a positive impact on the performance of the BPNN. The BPNN trained with the CopulaGAN2 (at Selangor River) and the TVAE2 (at Skudai River) yielded more accurate estimations compared to those derived from the actual and smaller datasets. Highlights: Data were insufficient to train machine learning model well in developing regions. Synthetic data methods can overcome the data scarcity issue in Malaysia. CopulaGAN and TVAE outperformed other methods at Selangor River and Skudai River. BPNN trained with synthetic datasets estimated WQI with higher accuracy. [ABSTRACT FROM AUTHOR]

Prediction of reference evapotranspiration (ET 0) remains a challenge, especially with forward multi-step forecasting. The bottleneck facing current research is the limitation of the span of the forecasting time horizons, which can be rather disappointing, especially when long-term forecasting is desired. In this study, an explainable model structure, represented by a one-dimensional convolutional neural network (CNN-1D) was compared to the long short-term memory network (LSTM) and gated recurrent unit network (GRU), both formulated with black-box model method. The comparison included the application of different forecasting strategies (iterated vs. multiple-input–multiple-output (MIMO)) and approaches (direct vs. indirect). This study was conducted at four stations scattered across the Peninsular Malaysia. From the results of this study, the explainable CNN-1D model generally performed poorer than its black-box counterparts at most of the stations. The type of model and its structure, forecasting strategy and approach formed a complex relationship to indicate that there is no one-for-all solution in the case of the long-term prediction of monthly mean ET 0. Despite that, the GRU-based models stood out as the most well-suited option for the task, with the MIMO forecasting strategy being favoured over the iterated strategy. At the four stations, the average mean absolute error (MAE), root mean square error (RMSE), mean percentage error (MAPE) and the Kling–Gupta efficiency (KGE) of the best GRU models were 0.182 mm/day, 0.260 mm/day, 4.972 % and 0.747, respectively. It was found that the prediction residual of the best GRU models did not possess a clear trend as the forecasting horizon was lengthened. The results implied that theoretically, the forecasting time horizon could be extended over to a longer temporal scale without any deterioration in the model performance. This finding is positive as it brings about the possibility of allocating the water budget with higher confidence. Nevertheless, the LSTM and GRU models developed in this study, were believed to have more tremendous potential if they were to be designed with purpose (such as the integration of optimisation algorithm), instead of being a mere black-box structure. • Current empirical/machine learning models cannot accurately forecast long-term ET 0. • Deep neural networks were trained using iterated and MIMO forecast strategies. • CNN-1D (explainable structure), LSTM, and GRU (black-box structure) were selected. • GRU MIMO,D and GRU MIMO,ID were found to have the best performance among the DNNs. • Prediction residual had a very weak correlation with the forecasting time horizon. [ABSTRACT FROM AUTHOR]

Over the past decade, there has been an increasing research on the use of machine learning tools for estimating reference crop evapotranspiration (ET o). However, due to the data-hungry nature of the machine learning models, all of these researches are not suitable for regions with limited data supply. This study aims to provide a breakthrough for the bottleneck through coupling of the inter-model ensemble with various data management schemes. The Bayesian modeling approach and a non-linear neural ensemble based inter-model ensemble (BMA-E and NNE-E) were developed locally with data from five different meteorological stations in the Peninsular Malaysia. The NNE-E was found to be highly robust spatially, whereby it can be used to estimate daily ET o accurately at other stations, even though with reduced input meteorological parameters. However, the performances of the locally trained models were found wanting and were fluctuating violently. This was resolved through creating a data pool that include the data from all stations and developing a universal NNE. By following the proposed scheme of things, the daily ET o can be easily estimated across the whole Peninsular Malaysia. This being, without the need for historical data and new models at estimation site. • Inter-model ensembles trained using exogenous data solve the data-hungry issues. • Bayesian modeling approach (BMA) and non-linear neural ensemble (NNE) were applied. • Black-box WOA-ELM-E model offered better spatial robustness than BMA-E model. • Use of global exogenous data pool further enhanced the ensemble's performance. [ABSTRACT FROM AUTHOR]

• MLP, SVM and ANFIS were hybridized using three different approaches. • Bootstrap aggregating did not improve the models due to good present data structure. • BMA only improved the models when its constituent models perform similarly. • WOA-ELM-E operated in black-box manner and overcome the constraints of base models. • GPI justified that WOA-ELM hybridization could develop the best hybrid models. Over recent years, many machine learning tools have been developed in the desire to replace conventional empirical models for reference evapotranspiration (ET 0) estimation that evidently needed tedious data collection and calibration. The feasibility of novel hybrid inter-model ensembles as a catalyst to further improve the estimation of ET 0 in a tropical climate region was investigated. The multilayer perceptron (MLP), support vector machine (SVM) and adaptive neuro-fuzzy inference system (ANFIS) were used as base model to estimate ET 0 at four meteorological stations of Peninsular Malaysia. Bootstrap aggregating, Bayesian model averaging (BMA) and extreme learning machine (ELM) based non-linear neural ensemble (NNE) were the three approaches used to hybridize the base models. The results showed that the bootstrap aggregating could not improve the performance of the base models by changing the data structure. The effectiveness of the BMA approach was constrained by the fact that it would favor the model with better accuracy. The whale optimization algorithm (WOA) optimized ELM operates in a black-box manner in order to compute the collective decisions of the MLP, SVM and ANFIS for estimating ET 0. The WOA-ELM, free to combine the favorable traits and features of the base models had resulted in a hybrid model with higher accuracy as well as generalizability. The global performance indicator (GPI) was used to compare and rank the models from all aspects. The GPI ranking system justified that the inter-model ensemble developed using the WOA-ELM-E (ensemble) was the best model scoring the highest value. Decision-makers are urged to take advantage of the hybrid models for more advantageous beneficial water management plans. [ABSTRACT FROM AUTHOR]

• Accurate estimation of ET 0 is needed although meteorological parameters are limited. • Simple empirical models gave estimation with insufficient accuracy and stability. • SVM has high computational cost when the problems have high dimensionality. • SVM mimics empirical models to reduce the dimensions and enhances ET 0 estimations. • The SVM-Rom model was recommended for tropical climate regions. In regions where economies that are highly dependent on agricultural activities, estimation of evapotranspiration is vital for scheduling and managing the water resources. In this study, the estimation of the reference evapotranspiration (ET 0) was performed for the west coast of Peninsular Malaysia, where oil palm is the major crop. Three stations, namely Station 48603 (Alor Setar), Station 48620 (Sitiawan) and Station 48650 (KLIA Sepang) were selected representing different regions along the west coast. Estimation of ET 0 using the popular and conventional Penman-Monteith (PM) model is data intensive. On the contrary, simpler empirical models such as the temperature-based Hargreaves-Samani (HS), mass transfer-based Romanenko and radiation-based Makkink models are less accurate, unstable and are greatly affected by local climate conditions. In order to solve the problems of the PM and other simple empirical models, the SVM was used to mimic the simple empirical models. The results of this study revealed that the HS and Makkink models were favorable for ET 0 estimation at the northern region, whereas for the central region, the Makkink model was not suitable. Simple averaging of the empirical models improved the estimations. However, the model was still deemed to be poor as the MARE ranged from 0.087 to 0.099, with the R2 still being possible to dip below 0.900. It was found that the SVM could greatly improve the ET 0 estimation. Unlike the empirical models, the regional factor had minimal effect on the ET 0 estimation; where the SVM-Mak model performed better than the SVM-HS and SVM-Rom models over the whole study area. The best performance was achieved by SVM-Mak model at Station 48620 (Sitiawan) where the MARE was only 0.020 when compared to the chosen standard PM model. It was found that the simple averaging of the SVM models could further boost the performance, but such an approach was non-rewarding. The SVM models were found to be better alternatives for the empirical models when using similar inputs. The study concluded that the SVM model which imitated the Makkink model was suitable for ET 0 estimation along the west coast of Peninsular Malaysia. [ABSTRACT FROM AUTHOR]