Algorithm 2 Edge-Aware Temperature Control via DDPG with Context-Aware Weights and Coordination

1:Input: Environment state $s=[T_{\mathrm{crate}},T_{\mathrm{ambient}},H,\mathrm{fruit}\mathrm{type}]$, spoilage model S, energy profile E   2:Initialize: Actor network $\mu \left(s\right|{\theta }^{\mu })$, critic network $Q(s,a|{\theta }^Q)$   3:Initialize: Target networks ${\mu }^{\prime }$ and $Q^{\prime }$ with ${\theta }^{{\mu }^{\prime }}\leftarrow {\theta }^{\mu }$, ${\theta }^{Q^{\prime }}\leftarrow {\theta }^Q$   4:Initialize: Replay buffer $\mathcal{R}$, noise process $\mathcal{N}$   5:Initialize: Shared intention buffer $\mathcal{B}\leftarrow \varnothing$, static coefficients ${\alpha }_1,{\alpha }_2$   6:for each episode do   7:   Receive initial global state $S_0=\{s_0^{\left(1\right)},s_0^{\left(2\right)},\dots \}$   8:   for each timestep t do   9:          for each agent i do 10:             Select action $a_t^{\left(i\right)}=\mu \left(s_t^{\left(i\right)}\right|{\theta }^{\mu })+{\mathcal{N}}_t$ 11:             Append $(s_t^{\left(i\right)},a_t^{\left(i\right)})$ to $\mathcal{B}$ 12:          end for 13:          Detect overlapping cooling requests in $\mathcal{B}$ (e.g., shared compressor or zone contention) 14:          if conflict detected then 15:             Apply coordination penalty $\rho$ or reschedule conflicting setpoints 16:          end if 17:          for each agent i do 18:             Apply action $a_t^{\left(i\right)}$, observe next state $s_{t+1}^{\left(i\right)}$, energy cost $E_t^{\left(i\right)}$, deviation $\mathrm{\Delta }{T}^{\left(i\right)}$, duration $\mathrm{\Delta }t$ 19:             Estimate spoilage risk ${\sigma }_t^{\left(i\right)}=S(s_t^{\left(i\right)},a_t^{\left(i\right)},\mathrm{\Delta }t)$ 20:             Define context vector: $ct{x}^{\left(i\right)}=[\mathrm{\Delta }{T}^{\left(i\right)},E_t^{\left(i\right)},{\sigma }_t^{\left(i\right)}]$ 21:              Compute dynamic weights: $${\omega }_j^{\left(i\right)}={\displaystyle \frac{{\alpha }_j·ct{x}_j^{\left(i\right)}}{{\sum }_{k=1}^3{\alpha }_k·ct{x}_k^{\left(i\right)}}},j=1,2,3$$ 22:              Compute reward: $$r_t^{\left(i\right)}=-({\omega }_1^{\left(i\right)}{E}_t^{\left(i\right)}+{\omega }_2^{\left(i\right)}{\sigma }_t^{\left(i\right)})-\rho$$ 23:             Store transition $(s_t^{\left(i\right)},a_t^{\left(i\right)},r_t^{\left(i\right)},s_{t+1}^{\left(i\right)})$ in $\mathcal{R}$ 24:          end for 25:          Sample mini-batch from $\mathcal{R}$ 26:           Update Critic using Bellman loss: $$L={\left(r+\gamma Q^{\prime }(s^{\prime },{\mu }^{\prime }\left(s^{\prime }\right))-Q(s,a)\right)}^2$$ 27:           Update actor via policy gradient: $${\nabla }_{{\theta }^{\mu }}J\approx {\displaystyle \frac{1}{N}}\sum {\nabla }_{a}Q(s,a){\nabla }_{{\theta }^{\mu }}\mu \left(s\right)$$ 28:           Soft update target networks: $${\theta }^{Q^{\prime }}\leftarrow \tau {\theta }^Q+(1-\tau ){\theta }^{Q^{\prime }},{\theta }^{{\mu }^{\prime }}\leftarrow \tau {\theta }^{\mu }+(1-\tau ){\theta }^{{\mu }^{\prime }}$$ 29:          Clear $\mathcal{B}\leftarrow \varnothing$ 30:      end for 31: end for 32:Output: Trained temperature control policy ${\mu }^{*}\left(s\right)$