From the flood situation in Thailand, it has had an impact on the livelihood of the population, especially in the agricultural sector, society, and the economy, particularly in Bangkok, which is an important economic city in Thailand. When floods occur in Bangkok, it can affect the overall economy of the country. Therefore, research and studies have been conducted to find sustainable solutions to the flood situation. In this regard, researchers have conducted a study to predict the water level in the Chao Phraya River using artificial neural network techniques, which involve synthesis and attribute screening using the Wrapper Method and variable reduction using the Backward Elimination method. The objective is to select data that has an impact on predicting the water level in the Chao Phraya River and analyze the number of nodes and hidden layers to achieve accurate predictions. The predicted outcome is the water quantity measured at the Lad Phrao canal level measurement point in the Chao Phraya River. The experiment involved building a neural network model using the Python programming language, utilizing the scikit-learn library, and filtering features through the SPSS program. The input data collected spans from January 1, 2560, to December 31, 2565, totaling 2,191 days, including data from reservoirs and dams (131 stations), rainfall data from weather stations (35 stations), sea level data from the Chulalongkorn University tide gauge station, and water level data from the Chao Phraya River. The data was divided into a training set, consisting of 1,753 days (80% of the total), and a test set, consisting of 438 days (20% of the total). The parameters of the neural network were adjusted to achieve optimal performance by varying the number of hidden layers from 1 to 6 and the number of nodes per layer from 500 to 2,000. The research found that there are 96 important features used for analysis, including 1) water level data from reservoirs and dams in the northern and central regions (22 stations), 2) rainfall data from the northern and central regions (61 stations), and 3) water level data from the connection point to the sea (1 station). The prediction was made by setting the number of hidden layers to 2 and each layer having 500 nodes. The model achieved a maximum prediction performance with an average deviation from the actual data of 13.13%.