超实用！Informer-LSTM时序预测+SHAP可解释性分析，手把手教你打造高精度模型

超实用Informer-LSTM时序预测SHAP可解释性分析手把手教你打造高精度模型精准捕捉长短期依赖让黑箱模型不再神秘在时间序列预测领域长序列预测一直是个挑战。今天我要向大家介绍一个强大的混合模型——Informer-LSTM并结合SHAP进行可解释性分析让你的预测模型既精准又透明一、为什么选择Informer-LSTM传统的时序预测模型在处理长序列时往往力不从心。RNN系列模型受限于梯度消失问题Transformer虽然强大但计算复杂度呈平方增长。Informer-LSTM混合模型巧妙地结合了两者优势Informer模块负责捕捉宏观的长期依赖LSTM模块精细处理局部的短期模式二、核心原理详解1. Informer模块高效处理长序列ProbSparse自注意力机制传统自注意力机制的复杂度是O(L²)当序列长度L10000时计算量巨大。ProbSparse注意力通过稀疏性评估只计算重要分数将复杂度优化到O(L log L)轻松处理超长序列。自注意力蒸馏在每层注意力模块后将序列长度裁剪一半像金字塔一样浓缩关键信息既减少计算量又提炼出最重要的长期依赖特征。生成式解码器传统模型需要逐步预测而Informer的解码器能一次性生成所有未来时间点的预测值极大提升效率。2. LSTM模块精准捕捉短期动态LSTM通过精巧的“门控”机制输入门、遗忘门、输出门克服了传统RNN的梯度消失问题特别擅长捕捉时间序列中复杂的短期动态和局部模式。三、实战完整实现流程步骤1环境配置与数据准备importnumpyasnpimportpandasaspdimporttorchimporttorch.nnasnnfromtorch.utils.dataimportDataLoader,TensorDatasetfromsklearn.preprocessingimportMinMaxScalerfromsklearn.metricsimportmean_squared_error,mean_absolute_error,r2_scoreimportmatplotlib.pyplotaspltimportshap# 设置随机种子torch.manual_seed(42)np.random.seed(42)步骤2构建Informer-LSTM模型classProbSparseAttention(nn.Module):ProbSparse自注意力机制def__init__(self,d_model,n_heads):super().__init__()self.d_modeld_model self.n_headsn_heads self.scale(d_model//n_heads)**-0.5defforward(self,Q,K,V):# 简化的ProbSparse实现batch_size,seq_len,_Q.shape# 采样部分关键查询sample_lenmin(seq_len,int(np.log(seq_len)))idxtorch.randperm(seq_len)[:sample_len]Q_sampleQ[:,idx,:]# 计算稀疏注意力分数scorestorch.matmul(Q_sample,K.transpose(-2,-1))*self.scale attntorch.softmax(scores,dim-1)# 只保留重要部分outputtorch.matmul(attn,V)returnoutputclassInformerLSTM(nn.Module):Informer-LSTM混合模型def__init__(self,input_size,hidden_size64,num_layers2,d_model128,n_heads8,output_size1):super().__init__()# Informer模块self.informer_attentionProbSparseAttention(d_model,n_heads)self.informer_fcnn.Linear(input_size,d_model)# LSTM模块self.lstmnn.LSTM(input_sized_model,hidden_sizehidden_size,num_layersnum_layers,batch_firstTrue,dropout0.2)# 输出层self.fcnn.Sequential(nn.Linear(hidden_size,32),nn.ReLU(),nn.Dropout(0.1),nn.Linear(32,output_size))defforward(self,x):# Informer路径x_informerself.informer_fc(x)attn_outputself.informer_attention(x_informer,x_informer,x_informer)# LSTM路径lstm_output,_self.lstm(attn_output)# 取最后一个时间步的输出outputself.fc(lstm_output[:,-1,:])returnoutput步骤3模型训练与评估deftrain_model(model,train_loader,val_loader,epochs100,lr0.001):训练模型并记录指标criterionnn.MSELoss()optimizertorch.optim.Adam(model.parameters(),lrlr)train_losses,val_losses,r2_scores[],[],[]best_val_lossfloat(inf)forepochinrange(epochs):# 训练阶段model.train()train_loss0forX_batch,y_batchintrain_loader:optimizer.zero_grad()y_predmodel(X_batch)losscriterion(y_pred,y_batch)loss.backward()optimizer.step()train_lossloss.item()# 验证阶段model.eval()val_loss0all_preds,all_true[],[]withtorch.no_grad():forX_batch,y_batchinval_loader:y_predmodel(X_batch)val_losscriterion(y_pred,y_batch).item()all_preds.extend(y_pred.numpy())all_true.extend(y_batch.numpy())# 计算R2r2r2_score(all_true,all_preds)train_losses.append(train_loss/len(train_loader))val_losses.append(val_loss/len(val_loader))r2_scores.append(r2)# 保存最佳模型ifval_lossbest_val_loss:best_val_lossval_loss torch.save(model.state_dict(),best_informer_lstm.pth)if(epoch1)%100:print(fEpoch [{epoch1}/{epochs}], fTrain Loss:{train_losses[-1]:.4f}, fVal Loss:{val_losses[-1]:.4f}, fR2:{r2:.4f})returntrain_losses,val_losses,r2_scores步骤4SHAP可解释性分析defshap_analysis(model,X_background,X_explain,feature_names):进行SHAP分析并可视化# 创建解释器explainershap.DeepExplainer(model,X_background)# 计算SHAP值shap_valuesexplainer.shap_values(X_explain)# 绘制各种可视化图表plt.figure(figsize(12,8))# 1. 蜂巢图shap.plots.beeswarm(shap_values,max_display10,showFalse)plt.title(SHAP蜂巢图 - 特征重要性分布)plt.tight_layout()plt.savefig(shap_beeswarm.png,dpi300,bbox_inchestight)plt.show()# 2. 瀑布图单个预测shap.plots.waterfall(shap_values[0],max_display10,showFalse)plt.title(SHAP瀑布图 - 单个预测解释)plt.tight_layout()plt.savefig(shap_waterfall.png,dpi300,bbox_inchestight)plt.show()# 3. 力图shap.initjs()shap.force_plot(explainer.expected_value,shap_values,X_explain,feature_namesfeature_names,matplotlibTrue,showFalse)plt.title(SHAP力图 - 特征贡献可视化)plt.tight_layout()plt.savefig(shap_force.png,dpi300,bbox_inchestight)plt.show()# 4. 决策图shap.decision_plot(explainer.expected_value,shap_values,feature_names,showFalse)plt.title(SHAP决策图 - 预测决策路径)plt.tight_layout()plt.savefig(shap_decision.png,dpi300,bbox_inchestight)plt.show()# 5. 热力图shap.plots.heatmap(shap_values,max_display10,showFalse)plt.title(SHAP热力图 - 特征影响模式)plt.tight_layout()plt.savefig(shap_heatmap.png,dpi300,bbox_inchestight)plt.show()# 6. 柱状图特征重要性shap.summary_plot(shap_values,X_explain,feature_namesfeature_names,plot_typebar,showFalse)plt.title(SHAP柱状图 - 特征重要性排序)plt.tight_layout()plt.savefig(shap_bar.png,dpi300,bbox_inchestight)plt.show()returnshap_values四、完整案例演示# 生成示例数据集defgenerate_sample_data(n_samples5000):生成示例时间序列数据tnp.arange(n_samples)# 趋势 季节 噪声trend0.01*t seasonalnp.sin(2*np.pi*t/50)0.5*np.sin(2*np.pi*t/20)noisenp.random.randn(n_samples)*0.1# 构建特征Xnp.column_stack([trendseasonalnoise,np.roll(trend,1)seasonal,np.sin(2*np.pi*t/30),np.cos(2*np.pi*t/30),np.random.randn(n_samples)*0.2])ytrendseasonalnoisereturnX,y# 准备数据lookback20# 时间步长X,ygenerate_sample_data()X_scaled,y_scaled,scaler_X,scaler_yprepare_data(X,y)# 创建序列数据X_seq,y_seqcreate_sequences(X_scaled,y_scaled,lookback)# 划分数据集train_sizeint(0.7*len(X_seq))val_sizeint(0.15*len(X_seq))test_sizelen(X_seq)-train_size-val_size train_datasetTensorDataset(X_seq[:train_size],y_seq[:train_size])val_datasetTensorDataset(X_seq[train_size:train_sizeval_size],y_seq[train_size:train_sizeval_size])test_datasetTensorDataset(X_seq[train_sizeval_size:],y_seq[train_sizeval_size:])# 训练模型modelInformerLSTM(input_sizeX.shape[1],output_size1)train_loaderDataLoader(train_dataset,batch_size64,shuffleTrue)val_loaderDataLoader(val_dataset,batch_size64)train_losses,val_losses,r2_scorestrain_model(model,train_loader,val_loader)# 评估模型model.load_state_dict(torch.load(best_informer_lstm.pth))test_loaderDataLoader(test_dataset,batch_size64)y_pred,y_trueevaluate_model(model,test_loader)# 计算评价指标msemean_squared_error(y_true,y_pred)rmsenp.sqrt(mse)maemean_absolute_error(y_true,y_pred)r2r2_score(y_true,y_pred)mapenp.mean(np.abs((y_true-y_pred)/y_true))*100print(fMSE:{mse:.4f})print(fRMSE:{rmse:.4f})print(fMAE:{mae:.4f})print(fR2:{r2:.4f})print(fMAPE:{mape:.2f}%)五、结果可视化defplot_results(y_true,y_pred,train_losses,val_losses,r2_scores):绘制所有结果图表fig,axesplt.subplots(2,3,figsize(15,10))# 损失曲线axes[0,0].plot(train_losses,labelTrain Loss)axes[0,0].plot(val_losses,labelVal Loss)axes[0,0].set_xlabel(Epoch)axes[0,0].set_ylabel(Loss)axes[0,0].set_title(训练和验证损失曲线)axes[0,0].legend()axes[0,0].grid(True)# R2曲线axes[0,1].plot(r2_scores)axes[0,1].set_xlabel(Epoch)axes[0,1].set_ylabel(R2 Score)axes[0,1].set_title(R2评分随epoch变化)axes[0,1].grid(True)# 回归散点图axes[0,2].scatter(y_true,y_pred,alpha0.5)axes[0,2].plot([y_true.min(),y_true.max()],[y_true.min(),y_true.max()],r--,lw2)axes[0,2].set_xlabel(True Values)axes[0,2].set_ylabel(Predictions)axes[0,2].set_title(f回归散点图 (R2{r2_score(y_true,y_pred):.3f}))axes[0,2].grid(True)# 残差图residualsy_true-y_pred axes[1,0].scatter(y_pred,residuals,alpha0.5)axes[1,0].axhline(y0,colorr,linestyle--)axes[1,0].set_xlabel(Predicted Values)axes[1,0].set_ylabel(Residuals)axes[1,0].set_title(残差图)axes[1,0].grid(True)# 时序对比图axes[1,1].plot(y_true[:200],labelTrue,alpha0.7)axes[1,1].plot(y_pred[:200],labelPredicted,alpha0.7)axes[1,1].set_xlabel(Time)axes[1,1].set_ylabel(Value)axes[1,1].set_title(时序预测对比图前200个样本)axes[1,1].legend()axes[1,1].grid(True)# 误差分布图axes[1,2].hist(residuals,bins50,edgecolorblack,alpha0.7)axes[1,2].set_xlabel(Residuals)axes[1,2].set_ylabel(Frequency)axes[1,2].set_title(残差分布直方图)axes[1,2].grid(True)plt.tight_layout()plt.savefig(all_results.png,dpi300,bbox_inchestight)plt.show()六、总结与建议Informer-LSTM混合模型在时序预测任务中表现出色优势既能捕捉长期依赖又能关注短期模式适用场景长序列预测、多变量时间序列、金融数据预测等可解释性通过SHAP分析让模型决策透明可信使用建议根据数据特征调整lookback参数尝试不同的隐藏层维度和层数使用早停法避免过拟合结合交叉验证选择最优超参数完整代码和数据集已准备好支持PyCharm和Jupyter Notebook使用包含详细的代码注释。欢迎留言交流本文代码使用PyTorch实现案例数据集可直接运行。