Prediction of Electrical Conductivity of Jet A-1 Fuel Using Multiple Regression Models

This paper presents the results of applying multiple linear regression to develop reliable models for predicting short-term (1–7 days) variation in the electrical conductivity of Jet A-1 aviation fuel, based on six years of monitoring data (2178 observations) of the fuel’s physicochemical properties. Predictor variables were selected based on paragenetic relationships with conductivity and included key physicochemical properties such as antistatic additive content, mercaptan sulfur content, smokeless flame height, calorific value, water content, aromatics, naphthalenes, crystallisation temperature, distillation temperatures (T10, T50, T90), colour, density, and microbial count. Lagged versions (0–6 days) were generated for each variable to capture temporal dependencies. Seven independent regression models were constructed to predict conductivity 1–7 days ahead using a direct multi-stage strategy. Model parameters were estimated using the QR-based least squares method, and predictor selection was performed through forward selection. Model performance was assessed using R2, MAPE, and 95% prediction intervals. Forecast accuracy decreased as the time horizon increased; however, even with a seven-day lead time, the model maintained a test-set MAPE below 15%, with a noticeably higher R2 value for 1–3 day forecast compared to naive models that predict the most recently available value as the future. The results show that short-term conductivity forecasting is feasible using interpretable linear models, and several physicochemical properties exhibit significant, lagged effects on conductivity variability.