PyTorch搭建LSTM实现多变量多步长时间序列预测（一）：直接多输出

I. 前言

在前面的两篇文章PyTorch搭建LSTM实现时间序列预测（负荷预测）和PyTorch搭建LSTM实现多变量时间序列预测（负荷预测）中，我们利用LSTM分别实现了单变量单步长时间序列预测和多变量单步长时间序列预测。

本篇文章主要考虑用PyTorch搭建LSTM实现多变量多步长时间序列预测。

系列文章：

II. 数据处理

数据集为某个地区某段时间内的电力负荷数据，除了负荷以外，还包括温度、湿度等信息。

本文中，我们根据前24个时刻的负荷以及该时刻的环境变量来预测接下来4个时刻的负荷（步长可调）。

任意输出其中一条数据：

(tensor([[0.5830, 1.0000, 0.9091, 0.6957, 0.8333, 0.4884, 0.5122],
        [0.6215, 1.0000, 0.9091, 0.7391, 0.8333, 0.4884, 0.5122],
        [0.5954, 1.0000, 0.9091, 0.7826, 0.8333, 0.4884, 0.5122],
        [0.5391, 1.0000, 0.9091, 0.8261, 0.8333, 0.4884, 0.5122],
        [0.5351, 1.0000, 0.9091, 0.8696, 0.8333, 0.4884, 0.5122],
        [0.5169, 1.0000, 0.9091, 0.9130, 0.8333, 0.4884, 0.5122],
        [0.4694, 1.0000, 0.9091, 0.9565, 0.8333, 0.4884, 0.5122],
        [0.4489, 1.0000, 0.9091, 1.0000, 0.8333, 0.4884, 0.5122],
        [0.4885, 1.0000, 0.9091, 0.0000, 1.0000, 0.3256, 0.3902],
        [0.4612, 1.0000, 0.9091, 0.0435, 1.0000, 0.3256, 0.3902],
        [0.4229, 1.0000, 0.9091, 0.0870, 1.0000, 0.3256, 0.3902],
        [0.4173, 1.0000, 0.9091, 0.1304, 1.0000, 0.3256, 0.3902],
        [0.4503, 1.0000, 0.9091, 0.1739, 1.0000, 0.3256, 0.3902],
        [0.4502, 1.0000, 0.9091, 0.2174, 1.0000, 0.3256, 0.3902],
        [0.5426, 1.0000, 0.9091, 0.2609, 1.0000, 0.3256, 0.3902],
        [0.5579, 1.0000, 0.9091, 0.3043, 1.0000, 0.3256, 0.3902],
        [0.6035, 1.0000, 0.9091, 0.3478, 1.0000, 0.3256, 0.3902],
        [0.6540, 1.0000, 0.9091, 0.3913, 1.0000, 0.3256, 0.3902],
        [0.6181, 1.0000, 0.9091, 0.4348, 1.0000, 0.3256, 0.3902],
        [0.6334, 1.0000, 0.9091, 0.4783, 1.0000, 0.3256, 0.3902],
        [0.6297, 1.0000, 0.9091, 0.5217, 1.0000, 0.3256, 0.3902],
        [0.5610, 1.0000, 0.9091, 0.5652, 1.0000, 0.3256, 0.3902],
        [0.5957, 1.0000, 0.9091, 0.6087, 1.0000, 0.3256, 0.3902],
        [0.6427, 1.0000, 0.9091, 0.6522, 1.0000, 0.3256, 0.3902]]), tensor([0.6360, 0.6996, 0.6889, 0.6434]))

数据格式为(X, Y)。其中X一共24行，表示前24个时刻的负荷值和该时刻的环境变量。Y一共四个值，表示需要预测的四个负荷值。需要注意的是，此时input_size=7，output_size=4。

III. LSTM模型

这里采用了深入理解PyTorch中LSTM的输入和输出（从input输入到Linear输出）中的模型：

class LSTM(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, output_size, batch_size):
        super().__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.output_size = output_size
        self.num_directions = 1 # 单向LSTM
        self.batch_size = batch_size
        self.lstm = nn.LSTM(self.input_size, self.hidden_size, self.num_layers, batch_first=True)
        self.linear = nn.Linear(self.hidden_size, self.output_size)

    def forward(self, input_seq):
        batch_size, seq_len = input_seq.shape[0], input_seq.shape[1]
        h_0 = torch.randn(self.num_directions * self.num_layers, self.batch_size, self.hidden_size).to(device)
        c_0 = torch.randn(self.num_directions * self.num_layers, self.batch_size, self.hidden_size).to(device)
        # output(batch_size, seq_len, num_directions * hidden_size)
        output, _ = self.lstm(input_seq, (h_0, c_0)) # output(5, 30, 64)
        pred = self.linear(output)  # (5, 30, 1)
        pred = pred[:, -1, :]  # (5, 1)
        return pred

IV. 训练和预测

训练和预测代码和前几篇都差不多，只是需要注意input_size和output_size的大小。

鉴于很多人一直在问代码，那这里重新贴一下好了，训练代码如下：

def train(args, Dtr, Val, path):
    if args.bidirectional:
        model = BiLSTM(args).to(device)
    else:
        model = LSTM(args).to(device)

    loss_function = nn.MSELoss().to(device)
    if args.optimizer == 'adam':
        optimizer = torch.optim.Adam(model.parameters(), lr=args.lr,
                                     weight_decay=args.weight_decay)
    else:
        optimizer = torch.optim.SGD(model.parameters(), lr=args.lr,
                                    momentum=0.9, weight_decay=args.weight_decay)
    scheduler = StepLR(optimizer, step_size=args.step_size, gamma=args.gamma)
    # training
    min_epochs = 10
    best_model = None
    min_val_loss = 5
    for epoch in tqdm(range(args.epochs)):
        train_loss = []
        for (seq, label) in Dtr:
            seq = seq.to(device)
            label = label.to(device)
            y_pred = model(seq)
            loss = loss_function(y_pred, label)
            train_loss.append(loss.item())
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

        scheduler.step()
        # validation
        val_loss = get_val_loss(args, model, Val)
        if epoch > min_epochs and val_loss < min_val_loss:
            min_val_loss = val_loss
            best_model = copy.deepcopy(model)

        print('epoch {:03d} train_loss {:.8f} val_loss {:.8f}'.format(epoch, np.mean(train_loss), val_loss))
        model.train()

    state = {'model': best_model.state_dict()}
    torch.save(state, path)

验证集上损失计算：

def get_val_loss(args, model, Val):
    model.eval()
    loss_function = nn.MSELoss().to(args.device)
    val_loss = []
    for (seq, label) in Val:
        with torch.no_grad():
            seq = seq.to(args.device)
            label = label.to(args.device)
            y_pred = model(seq)
            loss = loss_function(y_pred, label)
            val_loss.append(loss.item())

    return np.mean(val_loss)

测试：

def test(args, Dte, path, m, n):
    pred = []
    y = []
    print('loading models...')
    if args.bidirectional:
        model = BiLSTM(args).to(device)
    else:
        model = LSTM(args).to(device)
    model.load_state_dict(torch.load(path)['model'])
    model.eval()
    print('predicting...')
    for (seq, target) in tqdm(Dte):
        target = list(chain.from_iterable(target.data.tolist()))
        y.extend(target)
        seq = seq.to(device)
        with torch.no_grad():
            y_pred = model(seq)
            y_pred = list(chain.from_iterable(y_pred.data.tolist()))
            pred.extend(y_pred)

    y, pred = np.array(y), np.array(pred)
    y = (m - n) * y + n
    pred = (m - n) * pred + n
    # print(pred[-100:])
    print('mae:', get_mae(y, pred))
    print('mse', get_mse(y, pred))
    print('rmse:', get_rmse(y, pred))
    print('mape:', get_mape(y, pred))
    print('r2:', get_r2(y, pred))
    # plot
    plot(y, pred)

训练了50轮，预测接下来4个时刻的负荷值，MAPE为7.62%：
在这里插入图片描述

V. 源码及数据

后面将陆续公开~