Experimental Results and Artifacts for: A Causality-Preserving Framework for Multi-Horizon Forecasting from Degraded Wastewater Infrastructure Sensor Streams

In infrastructure sensor networks, predictive performance is often limited less by model capacity than by corrupted measurements, nonstationary operating regimes, and temporally invalid evaluation. This research presents a causality-preserving computational framework for multi-horizon forecasting from degraded wastewater-treatment sensor streams. The framework combines hybrid anomaly detection for mixed sensor failure modes, selective reconstruction, causal multi-resolution feature generation, and horizon-aware temporal validation. This validation explicitly separates preprocessing gains from model-family gains. The framework is demonstrated on real operational data from a resource-constrained municipal wastewater treatment plant (WWTP) serving 1,177 population equivalent (PE). The dataset comprises 19 sensor channels and 7,020 hourly observations after temporal consolidation. Four targets with contrasting dynamics are used to stress the method. These include inlet pH, inlet oxidation-reduction potential (ORP), effluent ammonium, and effluent nitrate. These variables span stable oscillatory behavior, abrupt electrochemical transitions, and episodic biologically coupled peaks. Results show that causally valid data preparation is the dominant contributor to predictive reliability. Model superiority depends on forecast horizon and signal type. The optimized Long Short-Term Memory (LSTM) architecture achieves strong performance for stable and biologically coupled targets. This includes R² = 0.927 for inlet pH and R² = 0.912 for effluent nitrate. It also maintains robust performance across extended forecast horizons for selected variables. The study is presented as a single-site demonstration of a general computational framework rather than as evidence of cross-plant generalization. Compatibility with future digital-twin integration is discussed at the architectural level. Multi-site transfer and closed-loop deployment remain subjects for future validation.