COMPARATIVE PERFORMANCE OF MLP AND LSTM SOFT-SENSOR MODELS FOR ESTIMATING CLASS A PAN WATER LEVEL FROM HIGH-RESOLUTION AUTOMATIC WEATHER STATION DATA

Abstract

Evaporation is a critical meteorological parameter for hydrology, agriculture, and climatology. However, its direct observation at the Indonesian Bureau of Meteorology, Climatology, and Geophysics (BMKG) stations heavily depends on manual hook-gauge readings and automatic water-level sensors, which are highly susceptible to data gaps during equipment malfunctions. This study developed and compared two soft-sensor models, a Multi-Layer Perceptron (MLP) and a Long Short-Term Memory (LSTM) network, to estimate the 5-minute water level in a Class A evaporation pan using Automatic Weather Station (AWS) variables. One year of 5-minute-resolution data (2025) from the Fatmawati Bengkulu Meteorological Station was utilized. The model inputs included air temperature, relative humidity, wind speed, solar radiation, barometric pressure, pan water temperature, and rainfall. Based on an Autocorrelation Function analysis, the temporal lookback window was fixed at 12 lags (60 minutes). The datasets were partitioned chronologically using a 70/15/15 split for training, validation, and testing, respectively. Evaluation metrics included RMSE, MAE, and R². The MLP model achieved a lower RMSE (2.9112 mm) and a higher R² (0.9742), which may suggest a greater robustness against large prediction errors and a stronger global fit. In contrast, the LSTM model yielded the lowest MAE (1.1048 mm), indicating a lower average absolute prediction error under typical conditions, though it exhibited higher sensitivity to abrupt environmental adjustments (RMSE: 3.6215 mm). These findings indicate that both architectures represent viable soft-sensor options. Ultimately, model selection should be determined by whether the operational system prioritizes overall error stability or lower average absolute error, thereby providing practical guidance for BMKG data-continuity protocols.