I. 前言

系列文章：

II. 多模型滚动预测

所谓多模型滚动预测：还是前10个预测后3个为例：首先需要按照多模型单步预测的方式训练3个模型，然后模型1利用[1…10]预测[11’]，然后模型2利用[2…10 11’]预测[12’]，最后由模型3利用[3…10 11’ 12’]预测[13’]。

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多模型单步预测一致：

# multiple model single step.
def nn_seq_mmss(B, pred_step_size):
    data = load_data()
    # 划分为训练集和测试
    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()
        # 需要针对每一个步骤返回多个seq
        seqs = [[] for i in range(pred_step_size)]
        for i in range(0, len(dataset) - 24 - pred_step_size, pred_step_size):
            train_seq = []

            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, ind in zip(range(i + 24, i + 24 + pred_step_size), range(pred_step_size)):
                # 形成多个label和seq
                train_label = [load[j]]
                seq = torch.FloatTensor(train_seq)
                train_label = torch.FloatTensor(train_label).view(-1)
                seqs[ind].append((seq, train_label))
        # 形成多个Dte
        res = []
        for seq in seqs:
            seq = MyDataset(seq)
            seq = DataLoader(dataset=seq, batch_size=batch_size, shuffle=False, num_workers=0, drop_last=True)
            res.append(seq)

        return res, [m, n]

    Dtrs, lis1 = process(train, B)
    Dtes, lis2 = process(test, B)

    return Dtrs, Dtes, lis1, lis2

简单来说，如果需要利用前24个值预测接下来12个值，那么我们需要生成12个数据集。

3.2 模型搭建

模型和之前的文章一致：

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
        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[0], input_seq[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))
        pred = self.linear(output)
        pred = pred[:, -1, :]
        return pred

3.3 模型训练/测试

模型训练与多模型单步预测一致：

Dtrs, Dtes, lis1, lis2 = load_data(args, flag, batch_size=args.batch_size)
for Dtr, path in zip(Dtrs, LSTM_PATHS):
    train(args, Dtr, path)

模型测试与单步滚动预测有些类似，但每一步都由不同的模型来进行预测：

# multiple model scrolling
def mms_rolling_test(args, Dte, lis, PATHS):
    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
    models = []
    for path in PATHS:
        if args.bidirectional:
            model = BiLSTM(input_size, hidden_size, num_layers, output_size, batch_size=args.batch_size).to(device)
        else:
            model = LSTM(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()
        models.append(model)

    Dte = [x for x in iter(Dte)]
    Dte = list_of_groups(Dte, args.pred_step_size)
    #
    for sub_item in tqdm(Dte):
        sub_pred = []
        for seq_idx, (seq, label) in enumerate(sub_item, 0):
            # 每个seq的batch都为1
            # 选定model
            model = models[seq_idx]
            label = list(chain.from_iterable(label.data.tolist()))
            y.extend(label)
            if seq_idx != 0:
                seq = seq.cpu().numpy().tolist()[0]
                # 如果当前预测长度已经大于seq，直接用pred的后几个将seq中每个数组的第一个数字替换掉
                if len(sub_pred) >= len(seq):
                    for t in range(len(seq)):
                        seq[t][0] = sub_pred[len(sub_pred) - len(seq) + t]
                else:
                    # 否则, seq的后几个用pred代替
                    for t in range(len(sub_pred)):
                        seq[len(seq) - len(sub_pred) + t][0] = sub_pred[t]
            else:
                # 第一个直接预测
                seq = seq.cpu().numpy().tolist()[0]
            # print(new_seq)
            seq = [seq]
            seq = torch.FloatTensor(seq)
            seq = MyDataset(seq)
            seq = DataLoader(dataset=seq, batch_size=1, shuffle=False, num_workers=0)
            # print(new_seq)
            seq = [x for x in iter(seq)][0]
            # print(new_seq)
            with torch.no_grad():
                seq = seq.to(device)
                y_pred = model(seq)
                y_pred = list(chain.from_iterable(y_pred.data.tolist()))
                # print(y_pred)
                sub_pred.extend(y_pred)

        pred.extend(sub_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，效果很差，MAPE为22.45%，还需要进一步完善。
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

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

文章出处登录后可见！

已经登录？立即刷新

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