代码：https://github.com/rajatsen91/deepglo

文章主要有三点创新：

• Leveled-Init TCN， 对 TCN 的改进，使用 LeveledInit 方法初始化 TCN 权重，使得 TCN 更好地处理不同尺度的时间序列；
• TCN-MF，用 TCN 来对多变量时间序列预测的矩阵分解法做正则化
• DeepGLO，把矩阵分解法 MF 得到的 包含全局信息的因子序列及其预测值 作为局部序列预测时的协变量。

LeveledInit TCN

从如下代码中可以看出对 TCN 权值初始化部分的改动

class Chomp1d(nn.Module):
    def __init__(self, chomp_size):
        super(Chomp1d, self).__init__()
        self.chomp_size = chomp_size

    def forward(self, x):
        return x[:, :, : -self.chomp_size].contiguous()


class TemporalBlock(nn.Module):
    def __init__(
        self,
        n_inputs,
        n_outputs,
        kernel_size,
        stride,
        dilation,
        padding,
        dropout=0.1,
        init=True,
    ):
        super(TemporalBlock, self).__init__()
        self.kernel_size = kernel_size
        self.conv1 = weight_norm(
            nn.Conv1d(
                n_inputs,
                n_outputs,
                kernel_size,
                stride=stride,
                padding=padding,
                dilation=dilation,
            )
        )
        self.chomp1 = Chomp1d(padding)
        self.relu1 = nn.ReLU()
        self.dropout1 = nn.Dropout(dropout)

        self.conv2 = weight_norm(
            nn.Conv1d(
                n_outputs,
                n_outputs,
                kernel_size,
                stride=stride,
                padding=padding,
                dilation=dilation,
            )
        )
        self.chomp2 = Chomp1d(padding)
        self.relu2 = nn.ReLU()
        self.dropout2 = nn.Dropout(dropout)

        self.net = nn.Sequential(
            self.conv1,
            self.chomp1,
            self.relu1,
            self.dropout1,
            self.conv2,
            self.chomp2,
            self.relu2,
            self.dropout2,
        )
        self.downsample = (
            nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None
        )
        self.init = init
        self.relu = nn.ReLU()
        self.init_weights()

    def init_weights(self):
        if self.init:
            nn.init.normal_(self.conv1.weight, std=1e-3)
            nn.init.normal_(self.conv2.weight, std=1e-3)

            self.conv1.weight[:, 0, :] += (
                1.0 / self.kernel_size
            )  ###new initialization scheme
            self.conv2.weight += 1.0 / self.kernel_size  ###new initialization scheme

            nn.init.normal_(self.conv1.bias, std=1e-6)
            nn.init.normal_(self.conv2.bias, std=1e-6)
        else:
            nn.init.xavier_uniform_(self.conv1.weight)
            nn.init.xavier_uniform_(self.conv2.weight)

        if self.downsample is not None:
            self.downsample.weight.data.normal_(0, 0.1)

    def forward(self, x):
        out = self.net(x)
        res = x if self.downsample is None else self.downsample(x)
        return self.relu(out + res)


class TemporalBlock_last(nn.Module):
    def __init__(
        self,
        n_inputs,
        n_outputs,
        kernel_size,
        stride,
        dilation,
        padding,
        dropout=0.2,
        init=True,
    ):
        super(TemporalBlock_last, self).__init__()
        self.kernel_size = kernel_size
        self.conv1 = weight_norm(
            nn.Conv1d(
                n_inputs,
                n_outputs,
                kernel_size,
                stride=stride,
                padding=padding,
                dilation=dilation,
            )
        )
        self.chomp1 = Chomp1d(padding)
        self.relu1 = nn.ReLU()
        self.dropout1 = nn.Dropout(dropout)

        self.conv2 = weight_norm(
            nn.Conv1d(
                n_outputs,
                n_outputs,
                kernel_size,
                stride=stride,
                padding=padding,
                dilation=dilation,
            )
        )
        self.chomp2 = Chomp1d(padding)
        self.relu2 = nn.ReLU()
        self.dropout2 = nn.Dropout(dropout)

        self.net = nn.Sequential(
            self.conv1,
            self.chomp1,
            self.dropout1,
            self.conv2,
            self.chomp2,
            self.dropout2,
        )
        self.downsample = (
            nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None
        )
        self.init = init
        self.relu = nn.ReLU()
        self.init_weights()

    def init_weights(self):
        if self.init:
            nn.init.normal_(self.conv1.weight, std=1e-3)
            nn.init.normal_(self.conv2.weight, std=1e-3)

            self.conv1.weight[:, 0, :] += (
                1.0 / self.kernel_size
            )  ###new initialization scheme
            self.conv2.weight += 1.0 / self.kernel_size  ###new initialization scheme

            nn.init.normal_(self.conv1.bias, std=1e-6)
            nn.init.normal_(self.conv2.bias, std=1e-6)
        else:
            nn.init.xavier_uniform_(self.conv1.weight)
            nn.init.xavier_uniform_(self.conv2.weight)

        if self.downsample is not None:
            self.downsample.weight.data.normal_(0, 0.1)

    def forward(self, x):
        out = self.net(x)
        res = x if self.downsample is None else self.downsample(x)
        return out + res


class TemporalConvNet(nn.Module):
    def __init__(self, num_inputs, num_channels, kernel_size=2, dropout=0.1, init=True):
        super(TemporalConvNet, self).__init__()
        layers = []
        self.num_channels = num_channels
        self.num_inputs = num_inputs
        self.kernel_size = kernel_size
        self.dropout = dropout
        num_levels = len(num_channels)
        for i in range(num_levels):
            dilation_size = 2 ** i
            in_channels = num_inputs if i == 0 else num_channels[i - 1]
            out_channels = num_channels[i]
            if i == num_levels - 1:
                layers += [
                    TemporalBlock_last(
                        in_channels,
                        out_channels,
                        kernel_size,
                        stride=1,
                        dilation=dilation_size,
                        padding=(kernel_size - 1) * dilation_size,
                        dropout=dropout,
                        init=init,
                    )
                ]
            else:
                layers += [
                    TemporalBlock(
                        in_channels,
                        out_channels,
                        kernel_size,
                        stride=1,
                        dilation=dilation_size,
                        padding=(kernel_size - 1) * dilation_size,
                        dropout=dropout,
                        init=init,
                    )
                ]

        self.network = nn.Sequential(*layers)

    def forward(self, x):
        return self.network(x)

TCN-MF

Temporal Convolution Network regularized Matrix Factorization

本质上和动态因子图是一回事：

一方面假设原始的多变量序列 $Y$ 是某些典型动态序列 $X$ 的线性组合:
$$Y=FX$$
另一方面，$X$ 的动态行为可以用 TCN 来建模：
$$X_t={\mathrm{T}}_X(X_{t-1})$$