feat: 新增了结合短期预测模型的全状态决策模型接口

models/uf-rl/rl_model/DQN/uf_decide/run_dqn_deicde_totalstate.py → models/uf-rl/rl_model/DQN/uf_decide/run_dqn_decide_totalstate.py

@@ -1,22 +1,32 @@
 """
 run_dqn_decide_totalstate.py
 
-全状态UF 超滤 DQN 决策主入口（Inference / Online Assist）
+全状态 UF 超滤 DQN 决策主入口（Inference / Online Assist）
 
 要求数据：
-1. UF1_prev_cycle.csv 示例数据，实际路径可更改，水厂上一周期完整数据，必须包含上一周期全部进水数据，且不能包含其他周期进水数据，分辨率1min，至少包含时间，进水流量，跨膜压差，温度
-2. UF1_init_cycle.csv 示例数据，实际路径可更改，水厂当前周期初始数据，建议在水厂当前周期进水稳定后传入并调用，程序计算该csv数据均值作为初始数据，eg：传入前10min数据，取周期前十分钟均值为初始值
+1. UF_prev_cycle.csv 示例数据，实际路径可更改，水厂上一周期完整数据，必须包含上一周期全部数据，不能包含其他周期数据，分辨率1min，至少包含时间，控制字/步序，进水流量，跨膜压差，温度
+2. UF_init_cycle.csv 示例数据，实际路径可更改，水厂当前周期初始实际数据，在水厂当前周期进水稳定后读取并传入，至少包含时间，控制字/步序，进水流量，跨膜压差，温度
+3. UF_predict_cycle.csv 示例数据，实际路径可更改，水厂当前预测数据，在调用决策模型前调用预测模型生成，包含未来20min的时间及渗透率
+
 职责：
-0. 根据水厂csv数据构造状态
 1. 构造物理世界（physics）
 2. 实例化决策器（UFDQNDecider）
-3. 构造当前工厂状态（observation）
-4. 调用模型给出策略建议
-5. 生成 PLC 下发指令（限幅 / 限速）
-6. 评估该指令在物理模型下的效果（只评估，不下发）
+3. 根据水厂csv数据构造当前工厂状态（observation）
+4. 调用模型生成模型指令
+5. 生成 PLC 下发指令
+6. 评估该指令在物理模型下的效果
 """
 
 from pathlib import Path
+from dataclasses import replace
+
+
+# ============================================================
+# 导入模块
+# ============================================================
+CURRENT_DIR = Path(__file__).resolve().parent
+
+UF_RL_ROOT = CURRENT_DIR.parents[2]     # uf_train  # uf-rl
 
 # ========== 参数 / 物理 ==========
 from env.uf_resistance_models_load import load_resistance_models
@@ -24,15 +34,14 @@ from env.uf_physics import UFPhysicsModel
 from env.env_params import UFState, UFActionSpec
 from env.env_config_loader import EnvConfigLoader, create_env_params_from_yaml
 
-
 # ========== 决策器 ==========
 from rl_model.DQN.uf_decide.dqn_decider import UFDQNDecider
+
 # ========== 决策状态构建器 ==========
 from rl_model.DQN.dqn_model.dqn_statebuilder import DQNStateBuilder
 
 
-
-def build_physics(IS_TIMES, phys_params):
+def build_physics(IS_TIMES, phys_params,state_bounds):
     """
     构造与训练一致的物理模型（只做一次）
     """
@@ -40,12 +49,75 @@ def build_physics(IS_TIMES, phys_params):
 
     physics = UFPhysicsModel(
         phys_params=phys_params,
+        state_bounds=state_bounds,
         resistance_model_fp=res_fp,
         resistance_model_bw=res_bw,
         IS_TIMES = IS_TIMES
     )
     return physics
 
+
+def check_state_bounds(current_state, state_bounds, unit_name):
+    """
+    检查当前状态是否在边界范围内
+
+    参数:
+        current_state: UFState对象，包含TMP, q_UF, temp
+        state_bounds: 状态边界对象
+        unit_name: 机组名称（如 "UF1"）
+
+    返回:
+        dict: 错误信息字典，格式 {"error_time": str, "error_feature": str}
+              如果没有错误，返回 None
+    """
+    from datetime import datetime
+
+    error_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+
+    # 检查各项参数是否在边界范围内
+    TMP0_min = state_bounds.TMP0_min
+    TMP0_max = state_bounds.TMP0_max
+    if not (TMP0_min <= current_state.TMP <= TMP0_max):
+        return {
+            "error_time": error_time,
+            "error_feature": f"{unit_name}Per"
+        }
+
+    nuK_min = state_bounds.nuK_min
+    nuK_max = state_bounds.nuK_max
+    if not (nuK_min <= current_state.nuK <= nuK_max):
+        return {
+            "error_time": error_time,
+            "error_feature": f"{unit_name}Per"
+        }
+
+    slope_min = state_bounds.slope_min
+    slope_max = state_bounds.slope_max
+    if not (slope_min <= current_state.slope <= slope_max):
+        return {
+            "error_time": error_time,
+            "error_feature": f"{unit_name}Per"
+        }
+
+    power_min = state_bounds.power_min
+    power_max = state_bounds.power_max
+    if not (power_min <= current_state.power <= power_max):
+        return {
+            "error_time": error_time,
+            "error_feature": f"{unit_name}Per"
+        }
+
+    ceb_removal_min = state_bounds.ceb_removal_min
+    ceb_removal_max = state_bounds.ceb_removal_max
+    if not (ceb_removal_min <= current_state.ceb_removal <= ceb_removal_max):
+        return {
+            "error_time": error_time,
+            "error_feature": f"{unit_name}Per"
+        }
+
+    return None
+
+
 def generate_plc_instructions(action_spec,current_L_s, current_t_bw_s, model_prev_L_s, model_prev_t_bw_s, model_L_s, model_t_bw_s):
     """
     根据工厂当前值、模型上一轮决策值和模型当前轮决策值，生成PLC指令。
@@ -105,16 +177,15 @@ def generate_plc_instructions(action_spec,current_L_s, current_t_bw_s, model_pre
     if model_prev_t_bw_s is not None and not (action_spec.t_bw_min_s <= model_prev_t_bw_s <= action_spec.t_bw_max_s):
         print(f"警告: 模型上一轮反洗时长 {model_prev_t_bw_s} 秒不在允许范围内 [{action_spec.t_bw_min_s}, {action_spec.t_bw_max_s}]")
 
-    # 模型当前轮决策值检查（错误）
+    # 模型当前轮决策值检查
     if model_L_s is None:
         raise ValueError("错误: 决策模型建议的过滤时长不能为None")
-    elif not (action_spec.L_min_s <= model_L_s <= action_spec.L_max_s):
-        raise ValueError(f"错误: 决策模型建议的过滤时长 {model_L_s} 秒不在允许范围内 [{action_spec.L_min_s}, {action_spec.L_max_s}]")
+    model_L_s = max(action_spec.L_min_s, min(model_L_s, action_spec.L_max_s))
 
     if model_t_bw_s is None:
         raise ValueError("错误: 决策模型建议的反洗时长不能为None")
-    elif not (action_spec.t_bw_min_s <= model_t_bw_s <= action_spec.t_bw_max_s):
-        raise ValueError(f"错误: 决策模型建议的反洗时长 {model_t_bw_s} 秒不在允许范围内 [{action_spec.t_bw_min_s}, {action_spec.t_bw_max_s}]")
+    model_t_bw_s = max(action_spec.t_bw_min_s, min(model_t_bw_s, action_spec.t_bw_max_s))
+
 
     print(f"过滤时长基准: {source_L}, 值: {effective_current_L}")
     print(f"反洗时长基准: {source_t_bw}, 值: {effective_current_t_bw}")
@@ -167,7 +238,8 @@ def calc_uf_cycle_metrics(current_state, max_tmp_during_filtration, min_tmp_duri
 
     # 获得模型模拟周期信息
     k_bw_per_ceb = info["k_bw_per_ceb"]
-    ton_water_energy_kWh_per_m3 = info["ton_water_energy_kWh_per_m3"]
+    refer_ton_water_energy = info["refer_ton_water_energy"]
+    ton_water_energy = info["ton_water_energy"]
     recovery = info["recovery"]
     daily_prod_time_h = info["daily_prod_time_h"]
 
@@ -183,7 +255,8 @@ def calc_uf_cycle_metrics(current_state, max_tmp_during_filtration, min_tmp_duri
 
     return {
         "k_bw_per_ceb": k_bw_per_ceb,
-        "ton_water_energy_kWh_per_m3": ton_water_energy_kWh_per_m3,
+        "refer_ton_water_energy": refer_ton_water_energy,
+        "ton_water_energy": ton_water_energy,
         "recovery": recovery,
         "daily_prod_time_h": daily_prod_time_h,
         "max_permeability": max_permeability
@@ -211,8 +284,8 @@ def run_dqn_decide(
         model_path=model_path,
         seed=0,
     )
-    # 模型决策（不推进真实环境）
 
+    # 模型决策
     decision = decider.decide(current_state)
     action_id = decision["action_id"]
     model_L_s = decision["L_s"]
@@ -226,22 +299,19 @@ def run_dqn_decide(
 # ==============================
 if __name__ == "__main__":
 
-    THIS_FILE = Path(__file__).resolve()
-    UF_RL_ROOT = THIS_FILE.parents[3]
-    CONFIG_PATH = UF_RL_ROOT / "xishan" / "uf_analyze_config.yaml"
-    ENV_CONFIG_PATH = UF_RL_ROOT / "xishan" / "env_config.yaml"
-    MODEL_PATH = UF_RL_ROOT / "xishan" / "48h_dqn_model.zip"
-    prev_cycle_csv = "test_online_datasets/UF1_prev_cycle.csv"
-    init_cycle_csv = "test_online_datasets/UF1_init_cycle.csv"
-    IS_TIMES = False # 新增指定变量，表示CEB间隔为时间控制/次数控制，T表示48次bw一次CEB，F表示48h一次CEB
-
-    # 构建强化学习状态
-    state_builder = DQNStateBuilder(config_path=CONFIG_PATH)
-    current_state: UFState = state_builder.build_from_csv_pair(
-        prev_cycle_csv=prev_cycle_csv,
-        init_cycle_csv=init_cycle_csv,
-    )
+    # ========== 模型及配置路径指定 ==========
+    IS_TIMES = False # 外部指定变量，表示CEB间隔为时间控制/次数控制，T表示48次bw一次CEB，F表示48h一次CEB
+    DATA_CONFIG_PATH = UF_RL_ROOT / "config_and_model" / "xishan" / "uf_analyze_config.yaml"
+    MODEL_PATH = UF_RL_ROOT / "config_and_model" / "xishan" / "48h_dqn_model.zip" # 需根据IS_TIMES变量值指定模型为48h_dqn_model.zip/48times_dqn_model.zip
+    ENV_CONFIG_PATH = UF_RL_ROOT / "config_and_model" / "xishan" / "env_config.yaml" # 环境配置路径
 
+    # ========== 外部调用输入 ==========
+    unit_name = "UF1"
+    prev_cycle_csv = CURRENT_DIR / "test_online_datasets" / "UF_prev_cycle.csv"
+    init_cycle_csv = CURRENT_DIR / "test_online_datasets" / "UF_init_cycle.csv"
+    predict_cycle_csv = CURRENT_DIR / "test_online_datasets" / "UF_predict_cycle.csv"
+
+    # ========== 模型及配置加载 ==========
     config_loader = EnvConfigLoader(ENV_CONFIG_PATH)
     config_loader.validate_config()
     config_loader.print_config_summary()
@@ -252,8 +322,26 @@ if __name__ == "__main__":
         reward_params,  # UFRewardParams
         state_bounds  # UFStateBounds
     ) = create_env_params_from_yaml(ENV_CONFIG_PATH)
-    physics = build_physics(IS_TIMES)
+    physics = build_physics(IS_TIMES, phys_params,state_bounds)
+
+    # ========== 调用模型生成模型指令 ==========
+    # 基于外部输入构建当前状态
+    state_builder = DQNStateBuilder(config_path=DATA_CONFIG_PATH)
+    current_state: UFState = state_builder.build_from_csv_pair(
+        unit_name,
+        uf_state_default,
+        state_bounds,
+        prev_cycle_csv=prev_cycle_csv,
+        init_cycle_csv=init_cycle_csv,
+        predict_cycle_csv=predict_cycle_csv
+    )
+
+    # 状态异常检查（仅检查，不中断，出现异常时后续归一化中将异常状态强制归一化至上下限）
+    error_result = check_state_bounds(current_state, state_bounds, unit_name)
+    if error_result:
+        print(f"错误发生时间: {error_result['error_time']}；错误特征量：{error_result['error_feature']}")
 
+    # 模型输出指令
     action_id, model_L_s, model_t_bw_s = run_dqn_decide(
         model_path=MODEL_PATH,
         physics=physics,
@@ -263,15 +351,15 @@ if __name__ == "__main__":
         current_state=current_state,
     ) # 环境实例化，模型加载等功能放在UFDQNDecider类中
 
+    # ========== 生成工厂下发指令 ==========
     current_L_s = 3800
     current_t_bw_s = 40
     model_prev_L_s = 4040
     model_prev_t_bw_s = 60
-    L_s, t_bw_s = generate_plc_instructions(current_L_s, current_t_bw_s, model_prev_L_s, model_prev_t_bw_s, model_L_s,
+    L_s, t_bw_s = generate_plc_instructions(action_spec, current_L_s, current_t_bw_s, model_prev_L_s, model_prev_t_bw_s, model_L_s,
                                             model_t_bw_s)  # 获取模型下发指令
 
-    L_s = 4100
-    t_bw_s = 96
+    # ========== 生成指令模拟执行结果 ==========
     max_tmp_during_filtration = 0.050176 # 新增工厂数据接口：周期最高/最低跨膜压差，无工厂数据接入时传入None，calc_uf_cycle_metrics()自动获取模拟周期中的跨膜压差最值
     min_tmp_during_filtration = 0.012496
     execution_result = calc_uf_cycle_metrics(current_state, max_tmp_during_filtration, min_tmp_during_filtration, L_s, t_bw_s)
@@ -280,7 +368,8 @@ if __name__ == "__main__":
     print(f"模型选择的L_s: {model_L_s} 秒, 模型选择的t_bw_s: {model_t_bw_s} 秒")
     print(f"指令下发的L_s: {L_s} 秒, 指令下发的t_bw_s: {t_bw_s} 秒")
     print(f"指令对应的反洗次数: {execution_result['k_bw_per_ceb']}")
-    print(f"指令对应的吨水电耗: {execution_result['ton_water_energy_kWh_per_m3']}")
+    print(f"指令对应的理论参考吨水电耗: {execution_result['refer_ton_water_energy']}")
+    print(f"指令对应的计算吨水电耗: {execution_result['ton_water_energy']}")
     print(f"指令对应的回收率: {execution_result['recovery']}")
     print(f"指令对应的日均产水时间: {execution_result['daily_prod_time_h']}")
     print(f"指令对应的最高渗透率: {execution_result['max_permeability']}")

models/uf-rl/rl_model/DQN/uf_decide/test_online_datasets/UF_predict_cycle.csv

@@ -0,0 +1,22 @@
+time,UF1Per
+2024/9/28 8:57,128.7106581
+2024/9/28 8:58,128.523688
+2024/9/28 8:59,128.7225444
+2024/9/28 9:00,128.7960679
+2024/9/28 9:01,128.3732001
+2024/9/28 9:02,128.3547406
+2024/9/28 9:03,128.2604308
+2024/9/28 9:04,128.1878302
+2024/9/28 9:05,128.2315896
+2024/9/28 9:06,128.0728985
+2024/9/28 9:07,127.9547203
+2024/9/28 9:08,127.6789771
+2024/9/28 9:09,127.9532075
+2024/9/28 9:10,129.3650869
+2024/9/28 9:11,128.9794077
+2024/9/28 9:12,129.6007399
+2024/9/28 9:13,129.3473523
+2024/9/28 9:14,129.2578081
+2024/9/28 9:15,129.3611611
+2024/9/28 9:16,129.3308854
+2024/9/28 9:17,128.870858