The Impact of Preprocessing Approaches on Neural Network Performance: A Case Study on Evaporation in Adana, a Mediterranean Climate

Abstract

The application of artificial intelligence (AI) technologies is quickly expanding in water management. Additionally, the artificial neural network methodology has an advantage over traditional statistical approaches in that it does not need assumptions about the distribution of data and variables. These methods can be used if there is a large enough data collection and criteria relevant to the nature of the problem. Preprocessing data before utilizing it improves the performance of the AI model. Evaporation matters in water management, agriculture processes and soil science. It is critical to ensure proper estimation of evaporation losses for effective water resource planning and management particularly in drought-prone areas such as Adana. Artificial intelligence approaches can be applied successfully in evaporation calculation. In this research, we used the Standard scaler, power transformer, robust scaler quantile transformer (Uniform) and quantile transformer (Normal), and min-max scaler preprocessing techniques to preprocess the multilayer perceptron neural network (MLPNN). We also trained the MLPNN using unprocessed data and compared it to the results of the preprocessed model. In the setup of the model, daily temperature, pressure, wind, sunny hours, and humidity parameters covering the years 2018-2021 were presented as input to the MLPNN model. Consequently, we pinpoint that all preprocessing approaches produce better outcomes than unscaled. Although all models produced statistically high accuracy predictions according to statistical criteria, the MLPNN model established without transformation (test phase: r2: 0.93, NSE : 0.927, SMAPE: 10.77, RMSE: 1.2, MAE: 0.9) exhibited the lowest accuracy. The evaporation prediction model that was developed using the MLPNN-based standard scalar optimization algorithm exhibited the highest level of accuracy  (test phase: r2: 0.978, NSE: 0.977, SMAPE: 5.93, RMSE: 0.68, MAE: 0.48). Power Transformer (test phase: r2: 0.978, NSE: 0.977, SMAPE: 5.81, RMSE: 0.67, MAE: 0.49) showed second-degree promising results. It can be concluded from these results that the estimation of meteorological variables requires the scaling and presentation of data in a uniform structure. Therefore, improving efficiency and productivity in water management or agricultural processes can be enhanced by making more accurate evaporation estimates.