""" PhyTopo-Net模型架构 """ import torch import torch.nn as nn import torch.nn.functional as F class ResBlock(nn.Module): """标准ResNet残差块""" def __init__(self, channels=64): super(ResBlock, self).__init__() self.conv1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1) self.bn1 = nn.BatchNorm2d(channels) self.conv2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1) self.bn2 = nn.BatchNorm2d(channels) def forward(self, x): residual = x out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) out += residual out = F.relu(out) return out class CrossAttentionFusion(nn.Module): """ 交叉注意力融合模块 Query: Meteorology features Key & Value: Topography features """ def __init__(self, channels=64): super(CrossAttentionFusion, self).__init__() self.channels = channels # Query projection (from Met features) self.query_conv = nn.Conv2d(channels, channels, kernel_size=1) # Key & Value projection (from Geo features) self.key_conv = nn.Conv2d(channels, channels, kernel_size=1) self.value_conv = nn.Conv2d(channels, channels, kernel_size=1) # Output projection self.out_conv = nn.Conv2d(channels, channels, kernel_size=1) self.softmax = nn.Softmax(dim=-1) def forward(self, feat_met, feat_geo): """ Parameters: ----------- feat_met : torch.Tensor 气象特征 (B, C, H, W) feat_geo : torch.Tensor 地形特征 (B, C, H, W) Returns: -------- fused_feat : torch.Tensor 融合后的特征 (B, C, H, W) """ B, C, H, W = feat_met.shape # Project to Query, Key, Value Q = self.query_conv(feat_met) # (B, C, H, W) K = self.key_conv(feat_geo) # (B, C, H, W) V = self.value_conv(feat_geo) # (B, C, H, W) # Reshape for attention computation Q = Q.view(B, C, H * W) # (B, C, HW) K = K.view(B, C, H * W) # (B, C, HW) V = V.view(B, C, H * W) # (B, C, HW) # Compute attention scores # Attention = Q @ K^T / sqrt(C) attention = torch.bmm(Q.transpose(1, 2), K) # (B, HW, HW) attention = attention / (C ** 0.5) attention = self.softmax(attention) # Apply attention to Value attended = torch.bmm(V, attention.transpose(1, 2)) # (B, C, HW) attended = attended.view(B, C, H, W) # (B, C, H, W) # Project output attended = self.out_conv(attended) # Residual connection: add back to Met features fused_feat = feat_met + attended return fused_feat class PhyTopoNet(nn.Module): """ PhyTopo-Net: 使用精细地形引导优化粗分辨率ERA5风场 residual: 若为 True,输出 = input + 网络预测的残差,即学习 WRF - ERA5 的修正量 若为 False(默认),输出 = 网络预测的全场,与阶段一保持一致,便于复现 """ def __init__(self, residual=False): super(PhyTopoNet, self).__init__() self.residual = residual # ========== Dual Encoders ========== # Meteorology Encoder self.met_encoder = nn.Sequential( nn.Conv2d(7, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True) ) # Topography Encoder self.geo_encoder = nn.Sequential( nn.Conv2d(5, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True) ) # ========== Cross-Attention Fusion ========== self.fusion = CrossAttentionFusion(channels=64) # ========== Backbone (ResNet Blocks) ========== self.backbone = nn.Sequential( ResBlock(64), ResBlock(64), ResBlock(64), ResBlock(64), ResBlock(64), ResBlock(64), ResBlock(64), ResBlock(64) ) # ========== Dual-Head Decoder ========== # Head Wind: Outputs U, V, T (3 channels, removed RAIN) self.head_wind = nn.Sequential( nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(64, 3, kernel_size=1) # Output: U, V, T ) # Head Pressure: Outputs SLP (1 channel) self.head_press = nn.Sequential( nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(64, 1, kernel_size=1) # Output: SLP ) def forward(self, x_met, x_geo): """ Forward pass Parameters: ----------- x_met : torch.Tensor 气象输入 (B, 7, 64, 64) - ERA5变量 x_geo : torch.Tensor 地形输入 (B, 5, 64, 64) - 静态地形变量 Returns: -------- output : torch.Tensor 输出 (B, 4, 64, 64) - [U, V, SLP, T] """ # Dual Encoders feat_met = self.met_encoder(x_met) # (B, 64, 64, 64) feat_geo = self.geo_encoder(x_geo) # (B, 64, 64, 64) # Cross-Attention Fusion fused_feat = self.fusion(feat_met, feat_geo) # (B, 64, 64, 64) # Backbone Processing features = self.backbone(fused_feat) # (B, 64, 64, 64) # Dual-Head Decoder out_wind = self.head_wind(features) # (B, 3, 64, 64) - [U, V, T] out_press = self.head_press(features) # (B, 1, 64, 64) - [SLP] # Reassemble output to match label order: [U, V, SLP, T] u = out_wind[:, 0:1, :, :] # (B, 1, 64, 64) v = out_wind[:, 1:2, :, :] # (B, 1, 64, 64) t = out_wind[:, 2:3, :, :] # (B, 1, 64, 64) # Concatenate: [U, V, SLP, T] raw_out = torch.cat([u, v, out_press, t], dim=1) # (B, 4, 64, 64) if self.residual: # 残差学习:output = input + residual,学习 WRF - ERA5 的修正量 final_out = x_met[:, :4] + raw_out else: final_out = raw_out return final_out