Research on Methods and Effects of Improving Data Quality in Smart Heating

Suboptimal telemetry data quality fundamentally degrades dispatch optimization and thermal load forecasting in smart district heating networks. Existing preprocessing routines rely heavily on isolated, unidimensional thresholds and bidirectional interpolation, routinely inducing high false-alarm rates during legitimate peak operations and causing acausal information leakage. To resolve these limitations, this study proposes an end-to-end data enhancement framework combining an Enhanced Isolation Forest with a strictly causal Long Short-Term Memory (LSTM) sequence imputation architecture. The anomaly detection module integrates Seasonal-Trend decomposition using Loess (STL) to eliminate diurnal cyclical masking, adopts an inverse-variance weighting scheme to prioritize discriminative variables, and implements Tikhonov-regularized Mahalanobis distance metric traversal to capture coupled thermodynamic covariance distortions. For sequential recovery, the causal LSTM network reconstructs unobserved states using solely historical antecedents, with mass flow rate algebraically recovered via thermal energy balance to preserve cross-parameter physical consistency. Validated on continuous hourly field SCADA observations across a complete heating season (N=2904), the proposed detection scheme achieves an F1-score of 94.55% with a low false positive rate of 2.45%, outperforming conventional isolation trees. In sequence reconstruction across a 672 h benchmark, the causal LSTM achieves normalized mean squared errors of 0.1196 for contiguous block voids and 0.1104 for single-point missing values, substantially surpassing classical Lagrange interpolation. Downstream deployment into a Bayesian-optimized Gated Recurrent Unit (GRU) load forecasting model demonstrates that this upstream data quality enhancement reduces the root mean squared error from 174.75 to 56.36 kWh/h (a 67.75% relative reduction), lowers the mean absolute percentage error from 12.17% to 4.05%, contracts error variance from 30,538.13 to 3240.7, and achieves a high goodness-of-fit (R2=0.9917). These findings provide an empirical bridge between upstream physics-consistent telemetry refinement and downstream predictive operational reliability in industrial thermal systems.