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')

load dataset from a csv file

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

fill missing values with the last known value

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

resample to hourly data

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

normalize the data

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

split into training and testing datasets

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

convert data into sequences of length seq_len+timestep_pred

seq_len = 24 * 7 # one week timestep_pred = 24 # predict the next 24 hours 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)

Define the LSTM model

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

define particle

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

define PSO algorithm

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):
        # Update velocity
        for j, param in enumerate(self.particles[i].position):
            self.particles[i].velocity[j] = self.momentum * self.particles[i].velocity[j] \
                                            + 2 * torch.rand_like(param) * (
                                                        self.particles[i].best_position[j] - param) \
                                            + 2 * torch.rand_like(param) * (global_best_position[j] - param)

            # Clamp velocity
            self.particles[i].velocity[j].clamp_(-1, 1)

            # Update position
            self.particles[i].position[j] += self.lr * self.particles[i].velocity[j]

            # Clamp position
            self.particles[i].position[j].clamp_(-1, 1)

        # Evaluate fitness of new position
        for j, param in enumerate(self.particles[i].position):
            self.model.state_dict()[f'fc.weight'] = self.particles[i].position[0]
            self.model.state_dict()[f'fc.bias'] = self.particles[i].position[1]
            self.model.state_dict()[f'lstm.weight_ih_l{j}'] = self.particles[i].position[2 * j + 2]
            self.model.state_dict()[f'lstm.weight_hh_l{j}'] = self.particles[i].position[2 * j + 3]
            self.model.state_dict()[f'lstm.bias_ih_l{j}'] = self.particles[i].position[4 * j + 4]
            self.model.state_dict()[f'lstm.bias_hh_l{j}'] = self.particles[i].position[4 * j + 5]

        loss = self.evaluate_fitness()

        # Update personal best
        if loss < self.particles[i].best_loss:
            self.particles[i].best_position = [param.data.detach().clone() for param in self.particles[i].position]
            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

Train the model using PSO

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())
    for param in model_param:
        with torch.no_grad():
            for param, particle_param in zip(model.parameters(), selected_particles[i].position):
                param.copy_(particle_param)
        # param.data.copy_(selected_particles[i].position)

    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()
            # Update personal best
            for j, param in enumerate(selected_particles[i].position):
                if j == 0:
                    new_param = model.state_dict()[f'fc.weight'].view(hidden_dim * seq_len, 1).clone()
                    new_param.copy_(model.state_dict()[f'fc.weight'])
                    param.data.copy_(new_param.view(-1))
                elif j == 1:
                    param.data.copy_(model.state_dict()[f'fc.bias'])
                elif j % 2 == 0 and j != 0:
                    param.data.copy_(model.state_dict()[f'lstm.weight_ih_l{j // 2}'])
                elif j % 2 == 1 and j != 1:
                    param.data.copy_(model.state_dict()[f'lstm.weight_hh_l{j // 2}'])
                elif j % 4 == 2:
                    param.data.copy_(model.state_dict()[f'lstm.bias_ih_l{j // 4}'])
                elif j % 4 == 3:
                    param.data.copy_(model.state_dict()[f'lstm.bias_hh_l{j // 4}'])

            selected_particles[i].best_position = [param.data.detach().clone() for param in
                                                   selected_particles[i].position]
            selected_particles[i].best_loss = optimizer_pso.evaluate_fitness()

            # Update global best
            if selected_particles[i].best_loss < global_best_loss:
                global_best_loss = selected_particles[i].best_loss
                global_best_position = [param.data.detach().clone() for param in selected_particles[i].best_position]

        # Update particle swarm
        optimizer_pso.step()

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