Modeling total dissolved gas supersaturation in high dam reservoir using advanced novel bidirectional best-first disjoint aggregating M5-rule based algorithms: An innovative framework for ungauged regions

•TDG predicted by novel IAER-M5-Rule, DA-M5-Rule, RS-M5-Rule, and WIHW-M5-Rule models.•A new bidirectional best-first algorithm was employed for feature selection.•Models were developed in one station, while their generalization power was tested by another station.•All input variables except discharge were effective on TDG prediction.•The DA-M5-Rule model outperforms other models during the testing and validation stages. Total Dissolved Gas (TDG) supersaturation is a critical ecological indicator for assessing water quality downstream of high dam reservoirs. TDG concentration exceeding 110 % can cause gas-bubble trauma in fish, leading to mortality and adversely affecting other aquatic organisms. We used hourly datasets of Water Temperature, Barometric Pressure, Spill from Dam, Sensor Depth, and Discharge to calibrate the Bidirectional Best First (BBF) algorithm for feature selection and ensemble-based models (hybrid of Iterative Absolute Error Regression (IAER), Disjoint Aggregating (DA), Random Subspace (RS), and Weighted Instance Handler Wrapper (WIHW) with M5-Rule) for TDG prediction. The TDG prediction capabilities of these models were compared to those of the Support Vector Regression model using statistical metrics. The study determined that Spill from Dam significantly influences TDG prediction at both stations (30.3 % and 31.9 %), while Barometric Pressure is the least effective variable (13.9 % and 15.6 %. According to the BBF technique, an optimal input scenario consists of Spill from Dam, Water Temperature, Barometric Pressure, and sensor depth. The DA-M5-Rule model with root mean square error (RMSE) = 4.00 %, and uncertainty coefficient with 95 % confidence level (U95%) = 11.00 had the highest predictive power followed by IAER-M5-Rule (RMSE=4.41 %, U95%=12.16), WIHW-M5-Rule (RMSE=4.42 %, U95%=12.17), RS-M5-Rule (RMSE=4.61 %, U95%=12.23) and SVR (RMSE=5.71 %, U95%=15.00) at the testing stage, while for generalization stage DA-M5-Rule model (RMSE=4.21 %, U95%=11.08) had the highest generalization power, followed by RS-M5-Rule (RMSE=4.22 %, U95%=10.90), IAER-M5-Rule (RMSE=4.30 %, U95%=11.40), WIHW-M5-Rule (RMSE=4.36 %, U95%=11.44), and SVR (RMSE=5.00 %, U95%=13.50). The relative deviation of the developed models varied between 0.20–0.78 % during the testing stage and 1.72–2.00 % during the validation stage. Developed models in the current study can be applied as a promising tool across the USA. In addition, these models can be used to precisely p