In grid-deficient environments, residential energy systems face severe carbon emission penalties due to mandatory reliance on diesel standby generators during supply interruptions. In Iraq, summer peak loads routinely exceed grid capacity, triggering prolonged generator operation and dramatically increasing household carbon footprints. This study presents a deep Q-network (DQN) reinforcement learning framework for intelligent battery energy storage system (BESS) scheduling, targeting carbon emissions reduction through strategic peak shaving. The DQN agent learns optimal battery dispatch strategies by internalizing diurnal patterns in load and solar generation through temporal state features, enabling anticipatory control without requiring explicit external forecasting models. The system is trained on one-year operational data from a representative Iraqi residential installation and evaluated over the critical summer period (122 days, 35.5% grid unavailability). The results demonstrate a 54.8% ${\mathrm{CO}}_2$ reduction (306.5 kg versus 677.4 kg baseline), a 25.5% reduction in generator runtime, and a 23.7% reduction in operating costs for the studied configuration. The learned policy approaches 89.6% of perfect-foresight MILP performance while executing 35,000 times faster. A reward function sensitivity analysis across five weighting schemes confirms that the 20:1 carbon-to-cost priority ratio optimally balances environmental and economic objectives. Ablation studies quantify the mechanism contributions: anticipatory pre-charging accounts for 58% of the total improvement, discharge optimization for 44%, and real-time PV coordination for 22%. These findings establish DQN-based BESS optimization as a practically deployable decarbonization approach for residential systems in grid-constrained developing regions.

Effective water management requires accurate forecasting of dam inflows, rainfall, and actual electricity demand. In this paper, we propose a framework for intelligent control of Haditha Dam gates using the improved Harris Hawks Multi-Objective algorithm (IMOHHO) with Long Short-Term Memory (LSTM). The study aims to develop a framework for scheduling the operation of the gates that strikes a balance between maximizing hydropower generation, maintaining storage levels close to the target value, reducing unused discharge, and optimizing the operations of the gates. This ensures that a minimum net head of a 147 m water level in the dam reservoir is maintained. The proposed model was applied to the Haditha Dam using operational gates and a 48-step time horizon with an interval of 30 min that could be changed as needed. In the IMOHHO-LSTM algorithm, each filtered solution represents an integrated timeline for the opening of the six gates throughout the runtime; the algorithm relies on the concept of non-dominant solutions and creates an archive of optimal solutions using the Pareto interface to achieve a balance between different operational objectives. The final results showed the potential of the proposed algorithm to ensure the efficient operation of the dam, with a total power generation of $\mathrm{10,434.14}$ MWh, while the average output power was $434.76$ MW. The maximum capacity reached 660 MW, which represents the maximum design capacity of a modern dam. Furthermore, the model kept the storage level within a very narrow range of 2199 to 2206 Mm³, centering around the target storage value of 2200 Mm³. This confirms the passing of the IMOHHO-LSTM algorithm in balancing energy production with maintaining the operational requirements of the dam lake’s water level.

Motor current signature analysis (MCSA) enables non-invasive condition monitoring of three-phase induction motors, but most published studies use short laboratory records with artificially seeded faults and report only a discrete fault label without any measure of confidence. This paper proposes a physics-informed temporal convolutional network (PI-TCN) that is trained and tested on 12 months of stator-current records measured on a 90 kW industrial induction motor in service. Fault labels were assigned retrospectively by aligning the recordings with documented maintenance interventions. From each 10 s current snapshot, 60 features are computed that describe time-domain statistics, inter-phase symmetry, harmonic distortion, the energy at theoretical fault frequencies, and the shape of the spectrum. Using a 24 h history of these features, the network classifies 10 motor conditions, estimates a continuous fault severity between 0 and 1, and provides prediction intervals for the severity. To avoid optimistic results, the data are divided by fault episode, with a temporal gap between training and test data. On the independent test set, the PI-TCN reaches an accuracy of $0.991\pm 0.002$ and a macro-averaged $F_1$ score of $0.898\pm 0.009$ over five training runs, and estimates the severity of faulty windows with a mean absolute error of $0.031\pm 0.004$. A five-member ensemble of the network gives prediction intervals whose empirical coverage matches the nominal level from 50% to 95% (expected calibration error 0.015), whereas split-conformal intervals under-cover at the low levels because calibration and test windows are not exchangeable under the temporal split. A label-perturbation study shows that a one-day error in the fault onset lowers the macro-averaged $F_1$ by up to 0.11 or more than doubles the severity error, depending on its direction. Because the method uses only the existing plant data logger, it can be deployed without additional sensors.