import os import random import numpy as np import pandas as pd

import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import torch.utils.data as Data

from sklearn.model_selection import train_test_split from tqdm import tqdm

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

从CSV文件加载数据集

df = pd.read_csv('Electric_Power_Consumption.csv', delimiter=';', parse_dates={'dt' : ['Date', 'Time']}, infer_datetime_format=True, na_values=['nan','?'], index_col='dt')

使用最后已知的值填充缺失值

df.fillna(method='ffill', inplace=True)

重采样为每小时数据

df = df.resample('H').mean()

归一化数据

df = (df - df.mean()) / df.std()

拆分为训练和测试数据集

train_df, test_df = train_test_split(df, test_size=0.2, shuffle=False)

将数据转换为长度为seq_len+timestep_pred的序列

seq_len = 24*7 # 一周 timestep_pred = 24 # 预测接下来的24小时 train_data = [] for i in range(seq_len, len(train_df) - timestep_pred): train_data.append(train_df.iloc[i-seq_len:i].values.tolist())

train_data = np.array(train_data) train_labels = train_df.iloc[seq_len+timestep_pred:].values

test_data = [] for i in range(seq_len, len(test_df) - timestep_pred): test_data.append(test_df.iloc[i-seq_len:i].values.tolist())

test_data = np.array(test_data) test_labels = test_df.iloc[seq_len+timestep_pred:].values

print('training dataset shapes: train_data: %s and train_labels: %s' % (train_data.shape, train_labels.shape)) print('testing dataset shapes: test_data: %s and test_labels: %s' % (test_data.shape, test_labels.shape))

train_y_ts = torch.from_numpy(train_labels).float().to(device) train_X_ts = torch.from_numpy(train_data).float().to(device)

test_y_ts = torch.from_numpy(test_labels).float().to(device) test_X_ts = torch.from_numpy(test_data).float().to(device)

train_set = Data.TensorDataset(train_X_ts, train_y_ts) test_set = Data.TensorDataset(test_X_ts, test_y_ts)

num_clients = 10 num_selected = 10 num_rounds = 29 epochs = 1 batch_size = 64

traindata_split = torch.utils.data.random_split(train_set, [int(train_labels.shape[0] / num_clients)] * (num_clients - 1) + [train_labels.shape[0] - int(train_labels.shape[0] / num_clients) * (num_clients - 1)])

train_loader = [torch.utils.data.DataLoader(x, batch_size=batch_size, shuffle=True) for x in traindata_split] test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=True)

定义LSTM模型

class LSTM(nn.Module): def init(self, input_dim, hidden_dim, seq_len, num_layers=2): super(LSTM, self).init() self.input_dim = input_dim self.hidden_dim = hidden_dim self.seq_len = seq_len self.num_layers = num_layers

    self.lstm = nn.LSTM(input_dim, hidden_dim, num_layers, batch_first=True)
    self.fc = nn.Linear(hidden_dim*seq_len, 1)

def forward(self, x):
    h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).to(device)
    c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).to(device)

    out, (hn, cn) = self.lstm(x, (h0,c0))
    out = out.reshape(out.size(0), -1)
    out = self.fc(out)

    return out

定义粒子

class Particle: def init(self, model): # self.position = model.get_params().detach().clone() self.position = model.parameters().next().data.detach().clone() self.velocity = torch.zeros_like(self.position) self.best_position = self.position.detach().clone() self.best_loss = float('inf')

定义PSO算法

class PSO: def init(self, model, criterion, lr=0.01, momentum=0.8, weight_decay=0.1): self.model = model self.criterion = criterion self.lr = lr self.momentum = momentum self.weight_decay = weight_decay self.particles = [Particle(model) for _ in range(num_selected)]

def step(self):
    for i in range(num_selected):
        # 更新速度
        self.particles[i].velocity = self.momentum * self.particles[i].velocity \
                                     + 2 * torch.rand_like(self.particles[i].position) * (
                                                 self.particles[i].best_position - self.particles[i].position) \
                                     + 2 * torch.rand_like(self.particles[i].position) * (
                                                 global_best_position - self.particles[i].position)

        # 限制速度
        self.particles[i].velocity.clamp_(-1, 1)

        # 更新位置
        self.particles[i].position += self.lr * self.particles[i].velocity

        # 限制位置
        self.particles[i].position.clamp_(-1, 1)

        # 评估新位置的适应度
        self.model.set_params(self.particles[i].position)
        # loss = self.evaluate_fitness()
        # model_param = iter(self.model.parameters())
        # for param in model_param:
        #     param.data.copy_(selected_particles[i].position)
        loss = self.evaluate_fitness()

        # 更新个人最佳
        if loss < self.particles[i].best_loss:
            self.particles[i].best_position = self.particles[i].position.detach().clone()
            self.particles[i].best_loss = loss

def evaluate_fitness(self):
    total_loss = 0.0
    with torch.no_grad():
        for data, target in test_loader:
            output = self.model(data)
            loss = self.criterion(output, target)
            total_loss += loss.item() * len(data)

    return total_loss / len(test_set)

def get_best(self):
    best_particle = min(self.particles, key=lambda x: x.best_loss)
    return best_particle.best_position, best_particle.best_loss

使用PSO训练模型

input_dim = train_data.shape[2] hidden_dim = 128 model = LSTM(input_dim, hidden_dim, seq_len).to(device) criterion = nn.MSELoss() optimizer_pso = PSO(model=model, criterion=criterion)

global_best_loss = float('inf') global_best_position = None

for r in range(num_rounds): selected_particles = random.sample(optimizer_pso.particles, num_selected)

for i in range(num_selected):
    # optimizer = optim.Adam(selected_particles[i].position, lr=0.01, weight_decay=0.1)
    optimizer = optim.Adam(model.parameters(), lr=0.01, weight_decay=0.1)
    # model.set_params(selected_particles[i].position)
    model_param = iter(model.parameters())
    with torch.no_grad():
        for param, particle_param in zip(model.parameters(), selected_particles[i].position):
            param.copy_(particle_param)

    for epoch in range(epochs):
        for data, target in tqdm(train_loader[i],desc='Training round %d/%d client %d/%d' % (r + 1, num_rounds, i + 1, num_clients)):
            optimizer.zero_grad()
            output = model(data)
            loss = criterion(output, target)
            loss.backward()
            # 更新个人最佳
            selected_particles[i].best_position = selected_particles[i].position.detach().clone()
            selected_particles[i].best_loss = optimizer_pso.evaluate_fitness()

            # 更新全局最佳
            if selected_particles[i].best_loss < global_best_loss:
                global_best_loss = selected_particles[i].best_loss
                global_best_position = selected_particles[i].best_position.detach().clone()

            optimizer.step()
            # 更新个人最佳
            selected_particles[i].best_position = selected_particles[i].position.detach().clone()
            selected_particles[i].best_loss = optimizer_pso.evaluate_fitness()

            # 更新全局最佳
            if selected_particles[i].best_loss < global_best_loss:
                global_best_loss = selected_particles[i].best_loss
                global_best_position = selected_particles[i].best_position.detach().clone()

        # 更新粒子群
        optimizer_pso.step()

        # 打印进度
        print('Round %d/%d, Global best loss: %.4f' % (r + 1, num_rounds, global_best_loss))