processes. The proposed approach not only enables accurate short-term P80 prediction, but also establishes a differentiable surrogate environment that directly supports the offline reinforcement learning formulation developed in this work, in which the trained LSTM serves as the transition model for policy optimization without requiring online plant experimentation. 3. METHODOLOGY This paper proposes a data-driven methodology for short-term P80 control in an industrial grinding circuit, based on multivariate time-series modeling using LSTM forecasting networks as a high-fidelity surrogate environment with offline Reinforcement Learning (RL) to enable predictive and P80 control. Four LSTM architectures are systematically compared (a baseline exogenous formulation and a feature-engineered variant incorporating metallurgical priors) and two autoregressive models (a formulation with lagged P80 feedback and a hybrid combining both strategies). Offline RL strategy is then evaluated following the process variable taxonomy of Hodouin (2011), this is a model-based differentiable actor-critic. 3.1. Data Preprocessing Data preprocessing was applied consistently across all four architectures. Raw 1-second sensor data were cleaned against operational limits (validated in collaboration with experienced metallurgists to retain only physically meaningful observations) and aggregated to 5-minute intervals via mean downsampling, preserving process dynamics while reducing computational burden. Temporal integrity was enforced through strict datetime index validation. A chronological 70/15/15 train/validation/test split was applied to preserve causal structure and prevent information leakage. Independent Minmax scalers were fitted exclusively on the training partition and subsequently applied to validation and test sets, mapping each feature to [0, 1]; separate scalers were retained for input variables and the target P80. All scaler objects are saved alongside the model weights to ensure consistent preprocessing at inference time in the RL environment 3.2. Sequence Construction All four LSTM variants employ a sliding-window encoder that transforms the univariateindexed time series into a supervised three-dimensional tensor. A lookback window of W = 6 timesteps (representing 30 minutes of historical context) is used to predict a forecast horizon of H = 2 timesteps (10 minutes ahead). This horizon was specifically designed to approximate ball mill residence time dynamics. The resulting tensor structures are X ∈ RN ×6×F and y ∈ RN ×2×1, where F denotes the number of input features (which varies by model configuration) and N = T − W − H + 1 represents the number of valid windows after boundary trimming. The trailing singleton dimension in y explicitly encodes the univariate nature of the target variable (P80), maintaining dimensional consistency with deep learning frameworks where the final axis corresponds to feature channels. Figure 3 illustrates the use of a sliding window approach to create input/output pairs for the LSTM, non-autoregressive variants exclude P80 from the input window; autoregressive variants include lagged P80 as an additional input covariate.
RkJQdWJsaXNoZXIy MTM0Mzk2