Accurately predicting changes in the fire environment helps to accurately grasp the development trend of the fire and ensure the safety of personnel. Due to the coexistence of multiple parameters of the fire scene environment, the complex coupling relationship, and the time series and nonlinearity, it is difficult to establish an accurate prediction model. Therefore, this paper proposes a long-term and short-term memory network model of self-attention mechanism based on the improved Harris Hawk algorithm, which realizes the accurate prediction of the fire scene environment data. Firstly, the logistic mapping strategy, cosine weighting factor, and Gaussian perturbation strategy are introduced into the Harris Hawk optimization algorithm to enrich its population diversity, balance its global exploration and local development capabilities, and improve its convergence accuracy. Then, the improved Harris Hawk optimization algorithm is used to optimize the hyperparameter in the self-attention mechanism short-term memory network model, and the fire environment is predicted based on the optimized parameters. The simulation results show that the self-attention mechanism based on the improved Harris Hawk optimization algorithm has better long-term memory network model fitting effect and higher prediction accuracy.