AI-driven multivariate time series forecasting of transmembrane pressure in ultrafiltration: A stationarity-aware approach

Accurate forecasting of transmembrane pressure (TMP) is essential for effective membrane fouling control and maintenance scheduling in ultrafiltration (UF) systems. This study proposes a stationarity-aware, artificial intelligence-assisted framework that enables long-horizon TMP prediction based on operational data. The framework incorporates autoregressive deep learning models—Long Short-Term Memory (LSTM) and Gated Recurrent Unit —combined with stationarity diagnostics using the Augmented Dickey-Fuller test and first-order differencing. Non-stationarity in process variables such as permeate flux and pressure was identified as a key challenge that degrades forecasting performance. Applying differencing to stabilize the input data led to substantial improvements in model accuracy. Specifically, the R 2 value of the LSTM model increased from −4.625 to 0.623 after preprocessing, representing an improvement of 5.25 points. In the final model, the GRU-based predictor achieved an R 2 value of 0.890, demonstrating high predictive accuracy. The models were also validated on an independent UF system operating under moderately elevated fouling conditions, demonstrating strong generalization and robustness. The proposed framework enables data-driven forecasting using only routine plant sensor data, without the need for additional instrumentation. This makes it well-suited for integration into existing monitoring systems. By improving TMP prediction accuracy and enabling proactive maintenance, the method offers practical value for long-term UF operation, particularly in reducing unnecessary cleaning cycles and enhancing operational stability. Moreover, the methodology is flexible and can be extended to other membrane-based processes facing similar fouling dynamics. Its low deployment cost and minimal infrastructure requirements make it an attractive option for digital upgrades in existing treatment facilities. • AI-based framework enhances TMP prediction in UF using stationarity-informed preprocessing. • First-order differencing improves model stability and reduces autocorrelation in raw data. • LSTM and GRU models predict long-horizon TMP using short-sequence operational inputs. • Forecasting supports proactive cleaning and reduces fouling risks in UF system operations. • Model is validated on independent plant data with variable fouling and operating conditions.