DualFlow/models/uf-rl/超滤训练源码/DQN_decide.py

import numpy as np
from stable_baselines3 import DQN
from UF_super_RL.DQN_env import UFSuperCycleEnv
from UF_super_RL.DQN_env import UFParams

# 模型路径
MODEL_PATH = "dqn_model.zip"

# 加载模型（只加载一次，提高效率）
model = DQN.load(MODEL_PATH)

def run_uf_DQN_decide(uf_params, TMP0_value: float):
    """
    单步决策函数：输入原始 TMP0，预测并执行动作

    参数:
        TMP0_value (float): 当前 TMP0 值（单位与环境一致）

    返回:
        dict: 包含模型选择的动作、动作参数、新状态、奖励等
    """
    # 1. 实例化环境
    base_params = uf_params
    env = UFSuperCycleEnv(base_params)

    # 2. 将输入的 TMP0 写入环境
    env.current_params.TMP0 = TMP0_value

    # 3. 获取归一化状态
    obs = env._get_obs().reshape(1, -1)

    # 4. 模型预测动作
    action, _ = model.predict(obs, deterministic=True)

    # 5. 解析动作对应的 L_s 和 t_bw_s
    L_s, t_bw_s = env._get_action_values(action[0])

    # 6. 在环境中执行该动作
    next_obs, reward, terminated, truncated, info = env.step(action[0])

    # 7. 整理结果
    result = {
        "action": int(action[0]),
        "L_s": float(L_s),
        "t_bw_s": float(t_bw_s),
        "next_obs": next_obs,
        "reward": reward,
        "terminated": terminated,
        "truncated": truncated,
        "info": info
    }

    # 8. 关闭环境
    env.close()

    return result

def generate_plc_instructions(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。

    新增功能：
    1. 处理None值情况：如果模型上一轮值为None，则使用工厂当前值；
       如果工厂当前值也为None，则返回None并提示错误。
    """
    # 参数配置保持不变
    params = UFParams(
        L_min_s=3600.0, L_max_s=6000.0, L_step_s=60.0,
        t_bw_min_s=40.0, t_bw_max_s=60.0, t_bw_step_s=5.0,
    )

    # 参数解包
    L_step_s = params.L_step_s
    t_bw_step_s = params.t_bw_step_s
    L_min_s = params.L_min_s
    L_max_s = params.L_max_s
    t_bw_min_s = params.t_bw_min_s
    t_bw_max_s = params.t_bw_max_s
    adjustment_threshold = 1.0

    # 处理None值情况
    if model_prev_L_s is None:
        if current_L_s is None:
            print("错误: 过滤时长的工厂当前值和模型上一轮值均为None")
            return None, None
        else:
            # 使用工厂当前值作为基准
            effective_current_L = current_L_s
            source_L = "工厂当前值(模型上一轮值为None)"
    else:
        # 模型上一轮值不为None，继续检查工厂当前值
        if current_L_s is None:
            effective_current_L = model_prev_L_s
            source_L = "模型上一轮值(工厂当前值为None)"
        else:
            effective_current_L = model_prev_L_s
            source_L = "模型上一轮值"

    # 对反洗时长进行同样的处理
    if model_prev_t_bw_s is None:
        if current_t_bw_s is None:
            print("错误: 反洗时长的工厂当前值和模型上一轮值均为None")
            return None, None
        else:
            effective_current_t_bw = current_t_bw_s
            source_t_bw = "工厂当前值(模型上一轮值为None)"
    else:
        if current_t_bw_s is None:
            effective_current_t_bw = model_prev_t_bw_s
            source_t_bw = "模型上一轮值(工厂当前值为None)"
        else:
            effective_current_t_bw = model_prev_t_bw_s
            source_t_bw = "模型上一轮值"

    # 检测所有输入值是否在规定范围内（只对非None值进行检查）
    # 工厂当前值检查（警告）
    if current_L_s is not None and not (L_min_s <= current_L_s <= L_max_s):
        print(f"警告: 当前过滤时长 {current_L_s} 秒不在允许范围内 [{L_min_s}, {L_max_s}]")
    if current_t_bw_s is not None and not (t_bw_min_s <= current_t_bw_s <= t_bw_max_s):
        print(f"警告: 当前反洗时长 {current_t_bw_s} 秒不在允许范围内 [{t_bw_min_s}, {t_bw_max_s}]")

    # 模型上一轮决策值检查（警告）
    if model_prev_L_s is not None and not (L_min_s <= model_prev_L_s <= L_max_s):
        print(f"警告: 模型上一轮过滤时长 {model_prev_L_s} 秒不在允许范围内 [{L_min_s}, {L_max_s}]")
    if model_prev_t_bw_s is not None and not (t_bw_min_s <= model_prev_t_bw_s <= 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 [{t_bw_min_s}, {t_bw_max_s}]")

    # 模型当前轮决策值检查（错误）
    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 (L_min_s <= model_L_s <= 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 [{L_min_s}, {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 (t_bw_min_s <= model_t_bw_s <= 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 [{t_bw_min_s}, {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}")

    # 使用选定的基准值进行计算调整
    L_diff = model_L_s - effective_current_L
    L_adjustment = 0
    if abs(L_diff) >= adjustment_threshold * L_step_s:
        if L_diff >= 0:
            L_adjustment = L_step_s
        else:
            L_adjustment = -L_step_s
    next_L_s = effective_current_L + L_adjustment

    t_bw_diff = model_t_bw_s - effective_current_t_bw
    t_bw_adjustment = 0
    if abs(t_bw_diff) >= adjustment_threshold * t_bw_step_s:
        if t_bw_diff >= 0:
            t_bw_adjustment = t_bw_step_s
        else:
            t_bw_adjustment = -t_bw_step_s
    next_t_bw_s = effective_current_t_bw + t_bw_adjustment

    return next_L_s, next_t_bw_s


from UF_super_RL.DQN_env import simulate_one_supercycle
def calc_uf_cycle_metrics(p, TMP0, max_tmp_during_filtration, min_tmp_during_filtration, L_s: float, t_bw_s: float):
    """
    计算 UF 超滤系统的核心性能指标

    参数:
        p (UFParams): UF 系统参数
        L_s (float): 单次过滤时间（秒）
        t_bw_s (float): 单次反洗时间（秒）

    返回:
        dict: {
            "k_bw_per_ceb": 小周期次数,
            "ton_water_energy_kWh_per_m3": 吨水电耗,
            "recovery": 回收率,
            "net_delivery_rate_m3ph": 净供水率 (m³/h),
            "daily_prod_time_h": 日均产水时间 (小时/天)
            "max_permeability": 全周期最高渗透率(lmh/bar)
        }
    """
    # 将跨膜压差写入参数
    p.TMP0 = TMP0

    # 模拟该参数下的超级周期
    feasible, info = simulate_one_supercycle(p, L_s, t_bw_s)

    # 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"]
    recovery = info["recovery"]
    net_delivery_rate_m3ph = info["net_delivery_rate_m3ph"]
    daily_prod_time_h = info["daily_prod_time_h"]

    # 获得模型模拟周期内最高跨膜压差/最低跨膜压差
    if max_tmp_during_filtration is None:
        max_tmp_during_filtration = info["max_TMP_during_filtration"]
    if min_tmp_during_filtration is None:
        min_tmp_during_filtration = info["min_TMP_during_filtration"]

    # 计算最高渗透率
    max_permeability = 100 * p.q_UF / (128*40) / min_tmp_during_filtration


    return {
        "k_bw_per_ceb": k_bw_per_ceb,
        "ton_water_energy_kWh_per_m3": ton_water_energy_kWh_per_m3,
        "recovery": recovery,
        "net_delivery_rate_m3ph": net_delivery_rate_m3ph,
        "daily_prod_time_h": daily_prod_time_h,
        "max_permeability": max_permeability
    }


# ==============================
# 示例调用
# ==============================
if __name__ == "__main__":
    uf_params = UFParams()
    TMP0 = 0.03 # 原始 TMP0
    model_decide_result = run_uf_DQN_decide(uf_params, TMP0) # 调用模型获得动作
    model_L_s = model_decide_result['L_s'] # 获得模型决策产水时长
    model_t_bw_s = model_decide_result['t_bw_s'] # 获得模型决策反洗时长

    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, model_t_bw_s) # 获取模型下发指令

    L_s = 4100
    t_bw_s = 96
    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(uf_params, TMP0, max_tmp_during_filtration, min_tmp_during_filtration, L_s, t_bw_s)
    print("\n===== 单步决策结果 =====")
    print(f"模型选择的动作: {model_decide_result['action']}")
    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']}")