Accurately predicting the temperature of nutrient solution is the key to regulating root zone temperature in Nutrient Film Technique (NFT), which is of great significance for crop growth. However, due to the temporal, nonlinear and multi coupling characteristics of nutrient solution temperature, it is difficult to achieve continuous and accurate prediction. This article proposes an improved genetic algorithm (IGA) to optimize a multivariable long short‑term memory (LSTM) neural network for predicting nutrient solution temperature. By introducing a sine function, the fixed crossover and mutation probabilities in the genetic algorithm were optimized. Pearson correlation analysis was used to obtain features with strong correlation. Simultaneously constructing a matrix of features and time steps, and inputting it into the network for temperature prediction, the model's coefficient of determination for predicting 20-60 min was 0.954-0.985, the root mean square error was 0.183 ℃-0.365 ℃, and the average absolute error was 0.165 ℃-0.311 ℃. And validated under different clarity indices KT, the results showed that the model had the best prediction effect on nutrient solution temperature when it was 0.5>KT≥0.2 (cloudy), and the model could meet the required prediction accuracy requirements for production under other KT, providing an important basis for precise and efficient temperature control in the root zone.