meeting

PPO.py

@@ -31,6 +31,7 @@ class ActorPPO(nn.Module):
             if isinstance(layer, nn.Linear):
                 nn.init.orthogonal_(layer.weight, 1.0)
                 nn.init.constant_(layer.bias, 0.0)
+
         self.net.apply(init_weights)
 
     def forward(self, state):
@@ -111,7 +112,8 @@ class AgentPPO:
         self.get_reward_sum = self.get_reward_sum_gae if if_use_gae else self.get_reward_sum_raw
 
         self.cri = self.ClassCri(net_dim, state_dim, action_dim, layer_norm).to(self.device)
-        self.act = self.ClassAct(net_dim, state_dim, action_dim, layer_norm).to(self.device) if self.ClassAct else self.cri
+        self.act = self.ClassAct(net_dim, state_dim, action_dim, layer_norm).to(
+            self.device) if self.ClassAct else self.cri
         self.cri_target = deepcopy(self.cri) if self.if_use_cri_target else self.cri
         self.act_target = deepcopy(self.act) if self.if_use_act_target else self.act
 
@@ -203,19 +205,7 @@ class AgentPPO:
         buf_advantage = buf_r_sum - (buf_mask * buf_value[:, 0])
         return buf_r_sum, buf_advantage
 
-    # def get_reward_sum_gae(self, buf_len, ten_reward, ten_mask, ten_value) -> (torch.Tensor, torch.Tensor):
+    import torch
-    #     buf_r_sum = torch.empty(buf_len, dtype=torch.float32, device=self.device)  # old policy value
-    #     buf_advantage = torch.empty(buf_len, dtype=torch.float32, device=self.device)  # advantage value
-    #
-    #     pre_r_sum = 0
-    #     pre_advantage = 0  # advantage value of previous step
-    #     for i in range(buf_len - 1, -1, -1):
-    #         buf_r_sum[i] = ten_reward[i] + ten_mask[i] * gamma * pre_r_sum
-    #         pre_r_sum = buf_r_sum[i]
-    #         delta = ten_reward[i] + ten_mask[i] * gamma * ten_value[i + 1] - ten_value[i]
-    #         buf_advantage[i] = delta + ten_mask[i] * gamma * self.lambda_gae_adv * pre_advantage
-    #         pre_advantage = buf_advantage[i]
-    #     return buf_r_sum, buf_advantage
 
     def get_reward_sum_gae(self, buf_len, ten_reward, ten_mask, ten_value) -> (torch.Tensor, torch.Tensor):
         buf_r_sum = torch.empty(buf_len, dtype=torch.float32, device=self.device)  # old policy value
@@ -224,12 +214,33 @@ class AgentPPO:
         pre_r_sum = 0.0
         pre_advantage = 0.0  # advantage value of previous step
         for i in range(buf_len - 1, -1, -1):
+            # Calculate TD residual delta
+            if i == buf_len - 1:
+                next_value = 0.0
+            else:
+                next_value = ten_value[i + 1]
+            delta = ten_reward[i] + ten_mask[i] * gamma * next_value - ten_value[i]
+
             buf_r_sum[i] = ten_reward[i] + ten_mask[i] * pre_r_sum
             pre_r_sum = buf_r_sum[i]
-            buf_advantage[i] = ten_reward[i] + ten_mask[i] * (pre_advantage - ten_value[i])  # fix a bug here
+            pre_advantage = delta + ten_mask[i] * gamma * self.lambda_gae_adv * pre_advantage
-            pre_advantage = ten_value[i] + buf_advantage[i] * self.lambda_gae_adv
+            buf_advantage[i] = pre_advantage
+
         return buf_r_sum, buf_advantage
 
+    # def get_reward_sum_gae(self, buf_len, ten_reward, ten_mask, ten_value) -> (torch.Tensor, torch.Tensor):
+    #     buf_r_sum = torch.empty(buf_len, dtype=torch.float32, device=self.device)  # old policy value
+    #     buf_advantage = torch.empty(buf_len, dtype=torch.float32, device=self.device)  # advantage value
+    #
+    #     pre_r_sum = 0.0
+    #     pre_advantage = 0.0  # advantage value of previous step
+    #     for i in range(buf_len - 1, -1, -1):
+    #         buf_r_sum[i] = ten_reward[i] + ten_mask[i] * pre_r_sum
+    #         pre_r_sum = buf_r_sum[i]
+    #         buf_advantage[i] = ten_reward[i] + ten_mask[i] * (pre_advantage - ten_value[i])  # fix a bug here
+    #         pre_advantage = ten_value[i] + buf_advantage[i] * self.lambda_gae_adv
+    #     return buf_r_sum, buf_advantage
+
     @staticmethod
     def optim_update(optimizer, objective):
         optimizer.zero_grad()

plotDRL.py

@@ -20,10 +20,9 @@ def plot_optimization_result(datasource, directory):  # data source is dataframe
 
     # 绘制步长成本 in ax[0]
     axs[0, 0].cla()
-    axs[0, 0].set_ylabel('Costs')
+    axs[0, 0].set_ylabel('Cost')
     axs[0, 0].set_xlabel('Time (h)')
     axs[0, 0].bar(T, datasource['step_cost'])
-    # axs[0,0].set_xticks([i for i in range(24)],[i for i in range(1,25)])
 
     # 绘制soc和价格 in ax[1]
     axs[0, 1].cla()
@@ -67,10 +66,9 @@ def plot_optimization_result(datasource, directory):  # data source is dataframe
     # 绘制净负载曲线
     axs[1, 0].plot(T, datasource['netload'], label='Netload', drawstyle='steps-mid', alpha=0.7)
     axs[1, 0].legend(loc='upper right', bbox_to_anchor=(1.4, 1), fontsize=12, frameon=False, labelspacing=0.3)
-    # axs[1,0].set_xticks([i for i in range(24)],[i for i in range(1,25)])
 
     fig.savefig(f"{directory}/rl.svg", format='svg', dpi=600, bbox_inches='tight')
-    print('optimization results have been ploted and saved')
+    print('rl figure have been ploted and saved')
 
 
 def plot_evaluation_information(datasource, directory):
@@ -91,11 +89,10 @@ def plot_evaluation_information(datasource, directory):
     # 绘制不平衡度 in axs[0]
     axs[0, 0].cla()
     axs[0, 0].set_xlabel('Time (h)')
-    axs[0, 0].set_ylabel('Unbalance of Generation and Load')
+    axs[0, 0].set_ylabel('Power (kWh)')
     axs[0, 0].bar(eval_data['time_step'], eval_data['unbalance'], label='Exchange with Grid', width=0.4)
     axs[0, 0].bar(eval_data['time_step'] + 0.4, eval_data['netload'], label='Netload', width=0.4)
     axs[0, 0].legend(loc='upper right', bbox_to_anchor=(1.45, 1), fontsize=12, frameon=False, labelspacing=0.5)
-    # axs[0,0].set_xticks([i for i in range(24)],[i for i in range(1,25)])
 
     # 绘制能源充/放电与价格关系图 in ax[1]
     axs[0, 1].cla()
@@ -140,10 +137,10 @@ def plot_evaluation_information(datasource, directory):
     # 绘制奖励 in axs[3]
     axs[1, 1].cla()
     axs[1, 1].set_xlabel('Time (h)')
-    axs[1, 1].set_ylabel('Costs')
+    axs[1, 1].set_ylabel('Cost')
     axs[1, 1].bar(eval_data['time_step'], eval_data['operation_cost'])
     fig.savefig(f"{directory}/gurobi.svg", format='svg', dpi=600, bbox_inches='tight')
-    print('evaluation figure have been ploted and saved')
+    print('gurobi figure have been ploted and saved')
 
 
 def make_dir(directory, feature_change):

tools.py

@@ -94,6 +94,7 @@ def optimization_base_result(env, month, day, initial_soc):
     cost_wind = gp.quicksum(wind[t] * wind_cofficient for t in range(period))
 
     m.setObjective((cost_gen + cost_battery + cost_import - cost_export + cost_solar + cost_wind), GRB.MINIMIZE)
+    m.setParam(GRB.Param.FuncNonlinear, 1)
     m.optimize()
 
     # 记录数据便于绘图