In order to analyze and predict the compressive strength of backfill under the influence of sand-cement ratio, slurry concentration and slag powder proportion, uniaxial compression test was carried out on 17 groups of backfill specimens by the response surface method. The influence of single factor and multi-factor interaction on the compressive strength of backfill at 3, 7 and 28 days was analyzed, and the material ratio was optimized. The filling strength prediction model based on residual connection long short-term memory (LSTM) network was established. The results show that the sand-cement ratio, slurry concentration and slag powder ratio have significant effects on the strength of backfill, and the interaction between sand-cement ratio and slag powder ratio has significant effects on the strength of backfill at day 7 and day 28. The interaction between slurry concentration and slag powder ratio has significant influence on 28-day strength of backfill. In order to meet the requirements of filling strength, the strength of backfill body after 28 days of model optimization is 3.04 MPa, and the relative error of repeated tests is 3.12%. Based on residual connection LSTM, a prediction model with sand-cement ratio, slurry concentration and slag powder ratio as input variables and backfill strength as output variables was constructed. The model root mean square error (RMSE) is 0.45, coefficient of determination (R²) is 0.999, and mean absolute error (MAE) is 0.054. The prediction accuracy is better than the traditional LSTM network model, and it can accurately predict the compressive strength of backfill.