PyTorch搭建LSTM实现多变量多步长时间序列预测（五）：seq2seq

I. 前言

系列文章：

1. 深入理解PyTorch中LSTM的输入和输出（从input输入到Linear输出）
2. PyTorch搭建LSTM实现时间序列预测（负荷预测）
3. PyTorch搭建LSTM实现多变量时间序列预测（负荷预测）
4. PyTorch搭建双向LSTM实现时间序列预测（负荷预测）
5. PyTorch搭建LSTM实现多变量多步长时间序列预测（一）：直接多输出
6. PyTorch搭建LSTM实现多变量多步长时间序列预测（二）：单步滚动预测
7. PyTorch搭建LSTM实现多变量多步长时间序列预测（三）：多模型单步预测
8. PyTorch搭建LSTM实现多变量多步长时间序列预测（四）：多模型滚动预测
9. PyTorch搭建LSTM实现多变量多步长时间序列预测（五）：seq2seq
10. PyTorch中实现LSTM多步长时间序列预测的几种方法总结（负荷预测）

II. seq2seq

seq2seq由两部分组成：Encoder和Decoder。seq2seq的输入是一个序列，输出也是一个序列，经常用于时间序列预测。关于seq2seq的具体原理可以参考：DL入门(3)：循环神经网络（RNN）。

III. 代码实现

3.1 数据处理

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12个时刻的负荷（步长pred_step_size可调）。

数据处理代码和前面的直接多输出预测一致：

# Multiple outputs data processing.
def nn_seq_mo(B, num):
    data = load_data()

    train = data[:int(len(data) * 0.7)]
    test = data[int(len(data) * 0.7):len(data)]

    def process(dataset, batch_size):
        load = dataset[dataset.columns[1]]
        load = load.tolist()
        m, n = np.max(load), np.min(load)
        load = (load - n) / (m - n)
        dataset = dataset.values.tolist()
        seq = []
        for i in range(0, len(dataset) - 24 - num, num):
            train_seq = []
            train_label = []

            for j in range(i, i + 24):
                x = [load[j]]
                for c in range(2, 8):
                    x.append(dataset[j][c])
                train_seq.append(x)

            for j in range(i + 24, i + 24 + num):
                train_label.append(load[j])

            train_seq = torch.FloatTensor(train_seq)
            train_label = torch.FloatTensor(train_label).view(-1)
            seq.append((train_seq, train_label))

        seq = MyDataset(seq)
        seq = DataLoader(dataset=seq, batch_size=batch_size, shuffle=False, num_workers=0, drop_last=True)

        return seq, [m, n]

    Dtr, lis1 = process(train, B)
    Dte, lis2 = process(test, B)

    return Dtr, Dte, lis1, lis2

3.2 模型搭建

模型搭建分为三个步骤：编码器、解码器以及seq2seq。

首先是Encoder：

class Encoder(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, batch_size):
        super().__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.num_directions = 1
        self.batch_size = batch_size
        self.lstm = nn.LSTM(self.input_size, self.hidden_size, self.num_layers, batch_first=True, bidirectional=False)

    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, batch_size, self.hidden_size).to(device)
        c_0 = torch.randn(self.num_directions * self.num_layers, batch_size, self.hidden_size).to(device)
        # print(input_seq.size())
        # output(batch_size, seq_len, num_directions * hidden_size)
        output, (h, c) = self.lstm(input_seq, (h_0, c_0))
        return h, c

一般来讲编码器采用的就是RNN网络，这里采用了LSTM将原始数据进行编码，然后将LSTM的最后的隐状态和单元状态返回。

接着是解码器Decoder：

class Decoder(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
        self.batch_size = batch_size
        self.lstm = nn.LSTM(self.input_size, self.hidden_size, self.num_layers, batch_first=True, bidirectional=False)
        self.linear = nn.Linear(self.hidden_size, self.output_size)

    def forward(self, input_seq, h, c):
        batch_size, seq_len = input_seq.shape[0], input_seq.shape[1]
        output, _ = self.lstm(input_seq, (h, c))
        pred = self.linear(output)  # pred()
        pred = pred[:, -1, :]
        return pred

编码器同样也由LSTM组成，不过解码器的初始的隐状态和单元状态是编码器的输出。

3.3 模型训练/测试

模型训练：

def seq2seq_train(args, Dtr, path):
    input_size, hidden_size, num_layers = args.input_size, args.hidden_size, args.num_layers
    output_size = args.output_size
    batch_size = args.batch_size
    model = Seq2Seq(input_size, hidden_size, num_layers, output_size, batch_size=batch_size).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
    loss = 0
    for i in tqdm(range(args.epochs)):
        cnt = 0
        for (seq, label) in Dtr:
            cnt += 1
            seq = seq.to(device)
            label = label.to(device)
            y_pred = model(seq)
            loss = loss_function(y_pred, label)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            # if cnt % 100 == 0:
            #     print('epoch', i, ':', cnt - 100, '~', cnt, loss.item())
        print('epoch', i, ':', loss.item())

        scheduler.step()
    # save
    state = {'model': model.state_dict(), 'optimizer': optimizer.state_dict()}
    torch.save(state, path)

模型测试：

def seq2seq_test(args, Dte, lis, path):
    # Dtr, Dte, lis1, lis2 = load_data(args, flag, args.batch_size)
    pred = []
    y = []
    print('loading model...')
    input_size, hidden_size, num_layers = args.input_size, args.hidden_size, args.num_layers
    output_size = args.output_size
    model = Seq2Seq(input_size, hidden_size, num_layers, output_size, batch_size=args.batch_size).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)
    m, n = lis[0], lis[1]
    y = (m - n) * y + n
    pred = (m - n) * pred + n
    print('mape:', get_mape(y, pred))
    # plot
    plot(y, pred)

多步预测的每一步，都需要用不同的模型来进行预测。在正式预测时，数据的batch_size需要设置为1。

3.4 实验结果

前24个预测未来12个，每个模型训练50轮，MAPE为9.09%，还需要进一步完善。

IV. 源码及数据

源码及数据我放在了GitHub上，下载时请随手给个follow和star，感谢！
LSTM-MultiStep-Forecasting