<p>Supported by multisource big data and cloud computing, artificial intelligence (AI) is transforming environmental monitoring and modeling. However, many high-performance AI models rely on high-performance computing, which consumes considerable electrical energy and may limit their suitability for long-term edge-based water-quality forecasting, motivating potentially low-power on-device computing approaches. This study proposes a memristor-informed synaptic neural network (SyNN) framework as a device-aware pre-hardware evaluation platform. In the proposed framework, synaptic weight updates are guided by experimentally measured pulse-driven long-term potentiation and long-term depression behaviors of an N-doped TaO_x memristor. The device exhibits robust, repeatable nonvolatile switching with stable resistance states during repeated cycling. Hydrological and meteorological observations were used to predict water-quality variables, including dissolved oxygen, total nitrogen, total organic carbon, total phosphorus, and chlorophyll-a. Prediction results obtained by the SyNN model showed target-dependent forecasting performance, with relatively reliable predictions for dissolved oxygen and total phosphorus, whereas chlorophyll-a remained more challenging. Additional sensitivity analyses were conducted to examine the effect of representative memristor-related non-idealities on model performance. SHapley Additive exPlanations (SHAP) analysis suggests that temperature-related inputs provide consistent baseline control, whereas flow and rainfall become important primarily during hydrological events. The proposed framework is expected to contribute to future on-device water-quality forecasting by reducing computational burden and providing a device-aware basis for low-power adaptive monitoring under changing river conditions.</p><p></p>