building-agents/genrtic_gpu_7.22.py

import json
import os
import time
import numpy as np
import pandas as pd
import torch
from concurrent.futures import ThreadPoolExecutor
from deap import base, creator, tools, algorithms


def fitness_torch(individuals, price, load, temperature, irradiance, wind_speed, prev_soc, prev_Pg1, prev_Pg2, prev_Pg3,
                  device):
    individuals_torch = torch.tensor(individuals, device=device)
    num = individuals_torch.shape[0]
    Ac, Ag1, Ag2, Ag3, Av = [individuals_torch[:, i * period:(i + 1) * period] for i in range(5)]
    # soc = torch.zeros((num, period), device=device)
    # Pg1, Pg2, Pg3 = [torch.zeros((num, period), device=device) for _ in range(3)]
    # Cb, Cg1, Cg2, Cg3, Cs, Cw, Rs, Cp, Pe, Ps, Ee, Es = [torch.zeros((num, period), device=device) for _ in range(12)]

    soc = torch.clamp(prev_soc + 0.2 * Ac * 0.9, 0.2, 0.8)
    Pg1 = torch.clamp(prev_Pg1 + 100 * Ag1, min=0, max=150)
    Pg2 = torch.clamp(prev_Pg2 + 100 * Ag2, min=0, max=375)
    Pg3 = torch.clamp(prev_Pg3 + 200 * Ag3, min=0, max=500)
    Pso = torch.clamp((0.2 * irradiance + 0.05 * temperature - 9.25) * (1 + Av), min=0)
    Pw = torch.where(
        (wind_speed >= 3) & (wind_speed < 8),
        wind_speed ** 3 * 172.2625 / 1000,
        torch.where(
            (wind_speed >= 8) & (wind_speed < 12),
            64 * 172.2625 / 125,
            torch.zeros_like(wind_speed)
        )
    )
    P = Ac + Pg1 + Pg2 + Pg3 + Pso + Pw

    Ee = torch.where(P >= load, P - load, torch.zeros_like(P))
    Es = torch.where(P < load, load - P, torch.zeros_like(P))

    Cb = 0.01 * Ac + 0.1 * soc
    Cg1 = 0.0034 * Pg1 ** 2 + 3 * Pg1 + 30
    Cg2 = 0.001 * Pg2 ** 2 + 10 * Pg2 + 40
    Cg3 = 0.001 * Pg3 ** 2 + 15 * Pg3 + 70
    Cs = 0.01 * Pso
    Cw = 0.01 * Pw
    Rs = 0.5 * price * Ee
    Cp = price * Es
    Pe = torch.where(Ee > 100, (Ee - 100) * 50, torch.zeros_like(Ee))
    Ps = torch.where(Es > 100, (Es - 100) * 50, torch.zeros_like(Es))

    total_cost = torch.sum(Cb + Cg1 + Cg2 + Cg3 + Cs + Cw + Pe + Ps - Rs + Cp, dim=1)
    reward = -total_cost / 1000

    return reward.cpu().numpy(), soc.cpu().numpy(), Pg1.cpu().numpy(), Pg2.cpu().numpy(), Pg3.cpu().numpy()


def save_decision_values(best_ind, period, index):
    decisions = {
        'Ac': best_ind[0],
        'Ag1': best_ind[1],
        'Ag2': best_ind[2],
        'Ag3': best_ind[3],
        'Av': best_ind[4]
    }
    with open(f'decision_values_{period}_index_{index}.json', 'w') as f:
        json.dump(decisions, f)


def save_progress(population, period):
    population_data = [ind.tolist() for ind in population]
    with open(f'population_{period}.json', 'w') as f:
        json.dump({'period': period, 'population': population_data}, f)


def load_progress():
    if os.path.exists('population_gen_499.json'):
        with open('population_gen_499.json', 'r') as f:
            data = json.load(f)
        return data['population'], data['period']
    return None, 0


def check_bounds(func):
    def wrapper(*args, **kwargs):
        offspring = func(*args, **kwargs)
        if offspring[0] is None or offspring[1] is None:
            print("Error: One of the offspring is None", offspring)
            raise ValueError("Offspring cannot be None.")
        for child in offspring:
            for i in range(len(child)):
                if child[i] < -1:
                    child[i] = -1
                elif child[i] > 1:
                    child[i] = 1
        return offspring
    return wrapper


def main():
    period = 8760
    NGEN = 500
    CXPB, MUTPB = 0.7, 0.2
    batch_size = 500

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    data = pd.read_csv('./data.csv')
    price, load, temperature, irradiance, wind_speed = [data[col].values for col in
                                                        ['price', 'load', 'temperature', 'irradiance', 'wind_speed']]

    creator.create("FitnessMin", base.Fitness, weights=(-1.0,))
    creator.create("Individual", list, fitness=creator.FitnessMin)

    toolbox = base.Toolbox()
    toolbox.register("attr_float", np.random.uniform, -1, 1)
    toolbox.register("individual", tools.initRepeat, creator.Individual, toolbox.attr_float, 5)
    toolbox.register("population", tools.initRepeat, list, toolbox.individual)

    toolbox.register("mate", check_bounds(tools.cxBlend), alpha=0.5)
    toolbox.register("mutate", check_bounds(tools.mutGaussian), mu=0, sigma=0.1, indpb=0.2)
    toolbox.register("select", tools.selTournament, tournsize=3)
    toolbox.register("evaluate", fitness_torch)

    population = load_progress()
    if population is None:
        population = toolbox.population(n=NGEN)
        print("Initial population:", len(population), "and first individual:", population[0])

    prev_soc = torch.tensor([0.4] * NGEN, device=device)
    prev_Pg1 = torch.zeros(NGEN, device=device)
    prev_Pg2 = torch.zeros(NGEN, device=device)
    prev_Pg3 = torch.zeros(NGEN, device=device)

    for index in range(period):
        for gen in range(NGEN):
            start_time = time.time()

            offspring = algorithms.varAnd(population, toolbox, cxpb=CXPB, mutpb=MUTPB)
            num_individuals = len(offspring)

            with ThreadPoolExecutor() as executor:
                futures = []
                for i in range(0, num_individuals, batch_size):
                    batch = offspring[i:i + batch_size]
                    individuals = [ind[:] for ind in batch]
                    futures.append(
                        executor.submit(toolbox.evaluate, individuals, price[index], load[index], temperature[index],
                                        irradiance[index], wind_speed[index], prev_soc, prev_Pg1, prev_Pg2, prev_Pg3,
                                        device)
                    )

                for future in futures:
                    fitnesses, socs, Pg1s, Pg2s, Pg3s = zip(*future.result())
                    for ind, fitness in zip(offspring, fitnesses):
                        ind.fitness.values = (fitness,)

                    prev_soc[:len(socs)] = torch.tensor(socs, device=device)
                    prev_Pg1[:len(Pg1s)] = torch.tensor(Pg1s, device=device)
                    prev_Pg2[:len(Pg2s)] = torch.tensor(Pg2s, device=device)
                    prev_Pg3[:len(Pg3s)] = torch.tensor(Pg3s, device=device)

            population = toolbox.select(offspring, k=len(population))
            print("Population after selection:", population[:5])
            end_time = time.time()
            print(f"第 {index + 1}小时完成花费 {end_time - start_time:.2f} 秒")

        best_ind = tools.selBest(population, 1)[0]
        print('最佳个体:', best_ind)
        print('适应度:', best_ind.fitness.values)
        save_decision_values(best_ind, period, index)
        save_progress(population, period)

        prev_soc.fill_(0.4)
        prev_Pg1.fill_(0)
        prev_Pg2.fill_(0)
        prev_Pg3.fill_(0)
        prev_soc[:len(best_ind)] = torch.tensor(best_ind[:period], device=device)
        prev_Pg1[:len(best_ind)] = torch.tensor(best_ind[period:2 * period], device=device)
        prev_Pg2[:len(best_ind)] = torch.tensor(best_ind[2 * period:3 * period], device=device)
        prev_Pg3[:len(best_ind)] = torch.tensor(best_ind[3 * period:4 * period], device=device)


if __name__ == "__main__":
    main()