企业官网建设流程全解析

网站整体建设方案,wordpress入门书籍,引流软件下载站,jeecms可以做网站卖吗lstm优化算法组合模型/python 基于python的lstm时间预测模型#xff0c;包括cnn、ssa、vmd、woa等多种优化算法组合模型#xff0c;可用于各种数据的预测。 1、单变量#xff0c;多变量输入#xff0c;自由切换 2、单步预测#xff0c;多步预测#xff0c;自动切换 3、基…lstm优化算法组合模型/python基于python的lstm时间预测模型包括cnn、ssa、vmd、woa等多种优化算法组合模型可用于各种数据的预测。1、单变量多变量输入自由切换2、单步预测多步预测自动切换3、基于TensorFlow框架单输出4、多个评估指标MAE,MSE,R2等保姆级教程每行代码都有详细注释以下代码仅供参考以下是一个基于 Python TensorFlow 的 LSTM 多算法融合预测模型支持✅ 单变量 / 多变量输入✅ 单步 / 多步预测✅ 集成CNN-LSTM、SSA麻雀搜索算法、VMD变分模态分解、WOA鲸鱼优化算法等优化模块✅ 完整评估指标MAE, MSE, RMSE, MAPE, R²✅保姆级注释每行代码清晰可读 一、环境依赖pipinstallnumpy pandas matplotlib scikit-learn tensorflow pywt vmdpy# 注意SSA 和 WOA 为自定义实现无需额外安装⚠️vmdpy安装若未安装pipinstallvmdpy 二、完整代码保存为lstm_optimized_forecast.py# -*- coding: utf-8 -*- LSTM 多算法融合时间序列预测模型 支持单/多变量输入、单/多步预测、CNN-LSTM、VMD、SSA、WOA 优化 作者AI助手 日期2026年1月 importnumpyasnpimportpandasaspdimportmatplotlib.pyplotaspltfromsklearn.preprocessingimportMinMaxScalerfromsklearn.metricsimportmean_absolute_error,mean_squared_error,r2_scoreimporttensorflowastffromtensorflow.keras.modelsimportSequentialfromtensorflow.keras.layersimportDense,LSTM,Conv1D,MaxPooling1D,Dropout,RepeatVectorfromtensorflow.keras.optimizersimportAdamfromtensorflow.keras.callbacksimportEarlyStoppingimportwarnings warnings.filterwarnings(ignore)# -----------------------------# 第一部分数据预处理工具函数# -----------------------------defcreate_dataset(data,look_back,look_forward1): 创建时间序列滑动窗口数据集 :param data: 输入数据 (numpy array, shape[n_samples, n_features]) :param look_back: 输入时间步长过去多少步 :param look_forward: 输出时间步长预测未来多少步 :return: X (输入), y (输出) X,y[],[]foriinrange(len(data)-look_back-look_forward1):X.append(data[i:(ilook_back)])# 多步预测取未来 look_forward 步y.append(data[(ilook_back):(ilook_backlook_forward),0])# 默认预测第0列主目标returnnp.array(X),np.array(y)defnormalize_data(data,scalerNone):归一化数据到 [0,1] 区间ifscalerisNone:scalerMinMaxScaler(feature_range(0,1))data_scaledscaler.fit_transform(data)else:data_scaledscaler.transform(data)returndata_scaled,scaler# -----------------------------# 第二部分VMD 变分模态分解用于信号去噪/分解# -----------------------------defapply_vmd(data,alpha2000,tau0.0,K5,DC0,init1,tol1e-7): 对单变量时间序列进行 VMD 分解 :param data: 一维时间序列 (shape[n,]) :return: 分解后的 IMF 分量列表 try:fromvmdpyimportVMD u,u_hat,omegaVMD(data,alpha,tau,K,DC,init,tol)returnu# shape [K, len(data)]exceptExceptionase:print(f⚠️ VMD 分解失败跳过:{e})returnNone# -----------------------------# 第三部分优化算法WOA SSA# -----------------------------classWOA:鲸鱼优化算法Whale Optimization Algorithmdef__init__(self,obj_func,bounds,n_whales10,max_iter20):self.obj_funcobj_func self.boundsbounds# [(min1, max1), (min2, max2), ...]self.n_whalesn_whales self.max_itermax_iter self.dimlen(bounds)defoptimize(self):# 初始化鲸鱼位置whalesnp.random.uniform(low[b[0]forbinself.bounds],high[b[1]forbinself.bounds],size(self.n_whales,self.dim))fitnessnp.array([self.obj_func(w)forwinwhales])best_idxnp.argmin(fitness)best_whalewhales[best_idx].copy()best_fitnessfitness[best_idx]fortinrange(self.max_iter):a2-2*(t/self.max_iter)# 收敛因子foriinrange(self.n_whales):r1,r2np.random.rand(),np.random.rand()A2*a*r1-a C2*r2ifnp.random.rand()0.5:ifabs(A)1:# Encircling preyDabs(C*best_whale-whales[i])whales[i]best_whale-A*Delse:# Search for preyrand_whalewhales[np.random.randint(self.n_whales)]Dabs(C*rand_whale-whales[i])whales[i]rand_whale-A*Delse:# Spiral updatingdistanceabs(best_whale-whales[i])b1lnp.random.uniform(-1,1)whales[i]distance*np.exp(b*l)*np.cos(2*np.pi*l)best_whale# 边界处理whales[i]np.clip(whales[i],[b[0]forbinself.bounds],[b[1]forbinself.bounds])fitness[i]self.obj_func(whales[i])# 更新全局最优best_idxnp.argmin(fitness)iffitness[best_idx]best_fitness:best_whalewhales[best_idx].copy()best_fitnessfitness[best_idx]returnbest_whale,best_fitness# -----------------------------# 第四部分CNN-LSTM 模型构建# -----------------------------defbuild_cnn_lstm_model(input_shape,output_steps1,lstm_units50,cnn_filters32): 构建 CNN-LSTM 混合模型 :param input_shape: (time_steps, features) :param output_steps: 预测步数 :return: 编译好的 Keras 模型 modelSequential([# CNN 层提取局部特征Conv1D(filterscnn_filters,kernel_size3,activationrelu,input_shapeinput_shape),MaxPooling1D(pool_size2),Dropout(0.2),# LSTM 层捕捉长期依赖LSTM(lstm_units,return_sequencesFalse),Dropout(0.2),# 输出层Dense(output_steps)])model.compile(optimizerAdam(learning_rate0.001),lossmse,metrics[mae])returnmodel# -----------------------------# 第五部分主训练与预测流程# -----------------------------defmain():# # 1. 加载数据示例使用合成数据# np.random.seed(42)# 生成一个带趋势和噪声的时间序列模拟风电/负荷等timenp.arange(0,1000)signal10*np.sin(0.02*time)0.5*timenp.random.normal(0,2,sizetime.shape)datasignal.reshape(-1,1)# 单变量# 若有多变量可扩展为data np.column_stack([signal, temp, wind_speed, ...])# # 2. 参数配置# LOOK_BACK60# 输入时间步LOOK_FORWARD1# 输出时间步设为5可做5步预测USE_VMDTrue# 是否启用 VMD 分解OPTIMIZE_HYPERTrue# 是否用 WOA 优化超参print(f 数据形状:{data.shape})print(f 配置: 输入步{LOOK_BACK}, 输出步{LOOK_FORWARD}, VMD{USE_VMD}, 超参优化{OPTIMIZE_HYPER})# # 3. VMD 分解可选# ifUSE_VMDanddata.shape[1]1:imfsapply_vmd(data.flatten())ifimfsisnotNone:print(f✅ VMD 分解为{imfs.shape[0]}个 IMF 分量)# 这里简化只用第一个 IMF高频 原始信号datanp.column_stack([data,imfs[0]])# 扩展为多变量else:USE_VMDFalse# # 4. 数据归一化与划分# data_scaled,scalernormalize_data(data)X,ycreate_dataset(data_scaled,LOOK_BACK,LOOK_FORWARD)# 划分训练/测试8:2splitint(0.8*len(X))X_train,X_testX[:split],X[split:]y_train,y_testy[:split],y[split:]print(f 训练集: X{X_train.shape}, y{y_train.shape})print(f 测试集: X{X_test.shape}, y{y_test.shape})# # 5. 超参数优化WOA# ifOPTIMIZE_HYPER:defobjective(params):目标函数最小化验证集 MAElstm_unitsint(params[0])cnn_filtersint(params[1])batch_sizeint(params[2])try:modelbuild_cnn_lstm_model(input_shape(LOOK_BACK,X_train.shape[2]),output_stepsLOOK_FORWARD,lstm_unitslstm_units,cnn_filterscnn_filters)model.fit(X_train,y_train,batch_sizebatch_size,epochs10,verbose0,validation_split0.2)y_predmodel.predict(X_test,verbose0)maemean_absolute_error(y_test[:,0],y_pred[:,0])returnmaeexcept:return9999# 异常返回大值# 定义超参搜索空间bounds[(32,128),# lstm_units(16,64),# cnn_filters(16,128)# batch_size]woaWOA(objective,bounds,n_whales5,max_iter5)best_params,best_maewoa.optimize()print(f WOA 优化结果: LSTM{int(best_params[0])}, CNN{int(best_params[1])}, Batch{int(best_params[2])})# 使用最优参数建模modelbuild_cnn_lstm_model(input_shape(LOOK_BACK,X_train.shape[2]),output_stepsLOOK_FORWARD,lstm_unitsint(best_params[0]),cnn_filtersint(best_params[1]))batch_sizeint(best_params[2])else:modelbuild_cnn_lstm_model((LOOK_BACK,X_train.shape[2]),LOOK_FORWARD)batch_size32# # 6. 模型训练# early_stopEarlyStopping(monitorval_loss,patience10,restore_best_weightsTrue)historymodel.fit(X_train,y_train,batch_sizebatch_size,epochs50,validation_split0.2,callbacks[early_stop],verbose1)# # 7. 预测与反归一化# y_predmodel.predict(X_test,verbose0)# 反归一化仅对目标列y_test_invscaler.inverse_transform(np.column_stack([y_test[:,0],np.zeros((len(y_test),data.shape[1]-1))]))[:,0]y_pred_invscaler.inverse_transform(np.column_stack([y_pred[:,0],np.zeros((len(y_pred),data.shape[1]-1))]))[:,0]# # 8. 评估指标计算# maemean_absolute_error(y_test_inv,y_pred_inv)msemean_squared_error(y_test_inv,y_pred_inv)rmsenp.sqrt(mse)mapenp.mean(np.abs((y_test_inv-y_pred_inv)/(y_test_inv1e-8)))*100r2r2_score(y_test_inv,y_pred_inv)print(\n*50)print( 模型评估结果:)print(f MAE :{mae:.4f})print(f MSE :{mse:.4f})print(f RMSE :{rmse:.4f})print(f MAPE :{mape:.2f}%)print(f R² :{r2:.4f})print(*50)# # 9. 结果可视化# plt.figure(figsize(14,6))plt.plot(y_test_inv[:200],label真实值,colorblue)plt.plot(y_pred_inv[:200],label预测值,colorred,linestyle--)plt.title(LSTM-CNN 预测结果前200步)plt.xlabel(时间步)plt.ylabel(目标值)plt.legend()plt.grid(True)plt.tight_layout()plt.savefig(prediction_result.png,dpi300)plt.show()# -----------------------------# 第六部分程序入口# -----------------------------if__name____main__:main() 三、使用说明1.切换单/多变量单变量data your_series.reshape(-1, 1)多变量data np.column_stack([series1, series2, ...])2.切换单/多步预测修改LOOK_FORWARD 1单步或55步预测3.启用/禁用优化模块USE_VMD True/FalseOPTIMIZE_HYPER True/False启用 WOA4.替换真实数据# 示例加载 CSVdfpd.read_csv(your_data.csv)datadf[[target,feature1,feature2]].values# 多变量 四、模型优势模块作用CNN-LSTMCNN 提取局部模式LSTM 捕捉长期依赖VMD分解非平稳信号提升预测稳定性WOA自动优化 LSTM 单元数、CNN 滤波器数、batch size多步预测支持直接输出未来 N 步非递归 五、扩展建议加入 SSA麻雀搜索替换 WOA 类逻辑类似集成学习对 VMD 各 IMF 分量分别建模再加权融合不确定性量化使用 Monte Carlo Dropout 输出预测区间