Under the background of low carbonization, more adjustable loads have been put into the power grid and are becoming an important part of cross-time load demand response and power coordination. It is necessary to identify them from ordinary loads to meet the need for power balance. In this paper, we propose a user-adjustable capability mining method based on the improved Particle Swarm Optimization Algorithm (PSO) and the Long Short-Term Memory (LSTM). It is an effective method for mining and evaluating the adjustable capability of loads. After load classification under LSTM, the improved PSO which uses the nonlinear variation weight and mutational operation is combined with LSTM to ensure the overload convergence and improve the load forecasting accuracy. The LSTM hyper-parameters are optimized by the improved PSO for the purpose to tackle the difficulty in finding appropriate LSTM parameters. After precisely classifying and predicting the loads, it is simple to figure out the adjustable loads and therefore has the user-adjustable capability in total.