435 Table 3 – LSTM Model Performance at Different Forecast Horizon Horizon = 1 Horizon = 2 Model R2 (%) MAE (%) RMSE (microns) R2 (%) MAE (%) RMSE (microns) Baseline 0.62 6.25 9.61 0.60 6.47 9.79 Feature-engineering 0.71 5.10 8.35 0.69 5.43 8.68 Autoregressive 0.94 2.08 3.53 0.90 2.88 4.82 Hybrid autoregressive + feature-engineering 0.95 2.13 3.56 0.90 2.91 4.82 Although the Hybrid model (Model 4) achieved a slightly higher coefficient of determination (R2 = 0.95) compared to the Autoregressive model (Model 3, R2 = 0.94), the performance difference is marginal and may not be statistically significant. For this reason, Model 3 will be used for RL agent. RL agent is represented by a Differentiable Actor-Critic architecture where the LSTM serves as a "world model", allowing for high-precision control. The model successfully navigated from an initial P80 deviation (approx. 183.45 µm) to the target zone of 160 µm. By iteration 300, it achieved a positive Mean Return of +1.9363, proving it can maintain the process within the desired ±10 µm as shown in Table 4. Table 4 – Results and metrics for Differentiable Actor-Critic architecture. Metric Initial Value (iter=1) Final Value (iter=2) Mean return –32.83 +1.94 Across Loss 32.83 –1.94 P80 ~23.4 µm < ±10 µm Verification N/A Gradient Flow Verified 5. CONCLUSIONS The four LSTM architectures evaluated in this study confirm that autoregressive feedback is the dominant factor governing short-horizon P80 prediction accuracy in closed grinding circuits, autoregressive LSTM successfully captures P80 dynamics and supports differentiable gradient flow, a key requirement for model-based RL. The introduction of lagged P80 (at the 5-minute horizon) as an additional input produced a step-change: the autoregressive (model 3) achieved R² = 0.94 and MAE = 2.08 µm, reducing absolute prediction error by 67 % relative to the baseline. The hybrid (model 4) which combines autoregressive feedback with metallurgically derived composite features, reached R² = 0.95 and MAE = 2.13 µm which represents a statistically marginal improvement of 0.01 in R² at the cost of eight additional input features. These results establish that temporal self-reference captures the dominant share of achievable predictive improvement, and that domain-engineered features provide diminishing returns once autoregressive structure is present. At the 10-minute horizon, the autoregressive (model 3) retained R² = 0.90 and MAE = 2.88 µm, confirming robust generalization across both forecast steps. The autoregressive model (Model 3) was selected as the surrogate environment for offline RL training on the basis of this analysis. Its lower input dimensionality (nine features versus seventeen for the hybrid) reduces the state space complexity for the agent without sacrificing predictive fidelity, as the nearidentical RMSE values at both horizons (3.53 µm vs. 3.56 µm at t+1; 4.82 µm vs. 4.82 µm at t+2) confirm. This selection reflects a broader design principle: model complexity should be justified by measurable performance gains, not by the availability of additional features
RkJQdWJsaXNoZXIy MTM0Mzk2