Policy-Conditioned Attention-Enhanced Temporal Convolutional Networks for Carbon Tax Price Forecasting

Forecasting administratively determined carbon-tax rates requires models capable of tracking persistent trajectories and discrete policy revisions. Building on the dilated causal temporal convolutional design of Chen et al., we combine a policy-conditioned input with a temporal attention readout and an explicit last-value residual path, and a matched 2×2 design separates architectural and policy-information contributions. Across 64 simulated trajectories with 112 monthly steps each, the proposed model achieves a one-step holdout MAE of 0.1278 and RMSE of 0.1605, and its attention-augmented readout lowers MAE by 4.70% relative to the policy-conditioned plain TCN (0.1341). Advance policy information has the larger effect: withholding it raises MAE to 0.3236, while on 37 policy-jump targets the proposed model attains MAE 0.2207 versus 0.2317 for the policy-conditioned plain TCN. These findings quantify the complementary roles of temporal attention and announcement-time policy schedules and provide a chronological benchmark for policy-aware carbon-tax forecasting.