python小白，没有看书，仅根据网上查到LSTM时序预测实例，编写了一个多维序列的预测程序（多维回归预测？）。在LSTM结构上不知是否正确，苦恼于如何根据实际需要构建复杂的LSTM 神经网络结构，希望大佬能给予建议。
环境：python3.5，tensorflow1.2.0，win10

import scipy.io as sio
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from sklearn import preprocessing
import os
import shutilos.chdir('D:\\xxxx')
folder = os.getcwd()+ '\\pythonLSTMtraining\\'if not os.path.exists(folder):os.makedirs(folder)
else:shutil.rmtree(folder, ignore_errors=True)os.makedirs(folder)#import the data 
d_i = sio.loadmat(os.getcwd()+'\\INPUT.mat')
data_in = d_i['INPUT'].T
scaler1 = preprocessing.StandardScaler().fit(data_in)
data_input = scaler1.transform(data_in)
d_o = sio.loadmat(os.getcwd()+'\\OUTPUT.mat')
data_out = d_o['OUTPUT'].T
scaler2 = preprocessing.StandardScaler().fit(data_out)
data_output = scaler2.transform(data_out)#parameters
time_step = 1 #序列段长度
lstm_layers = 2 #cell层数
batch_size = 100 #序列段批处理数目
input_size = data_input.shape[1] #输入维度
output_size = data_output.shape[1] #输出维度
rnn_unit = output_size #隐藏层节点数目
lr = 0.002 #学习率
per = 0.9 #划分比例
train_begin = 0
train_end = round(per * data_input.shape[0]) 
test_begin = train_end
#制作带时间步长的训练集
data_input_train = data_input[train_begin:train_end,:]
data_output_train = data_output[train_begin:train_end,:]
train_x,train_y=[],[]
for i in range(len(data_input_train)-time_step-1):x = data_input_train[i:i+time_step,:]y = data_output_train[i:i+time_step,:]train_x.append(x.tolist())train_y.append(y.tolist()) 
#制作带时间步长的测试集
data_input_test = data_input[test_begin:,:]
data_output_test = data_output[test_begin:,:]
test_x,test_y=[],[]
for i in range(len(data_input_test)-time_step-1):x = data_input_test[i:i+time_step,:]y = data_output_test[i:i+time_step,:]test_x.append(x.tolist())test_y.append(y.tolist()) #build the LSTM model
#输入层、输出层权重、偏置
X=tf.placeholder(tf.float32, shape=[None,time_step,input_size])#占位分配内存
Y=tf.placeholder(tf.float32, shape=[None,time_step,output_size])weights = {'in': tf.Variable(tf.random_normal([input_size, rnn_unit])),'out': tf.Variable(tf.random_normal([rnn_unit, output_size]))#2和1有区别吗？后改为输出维度，不知是否正确
}
biases = {'in': tf.Variable(tf.constant(0.1, shape=[rnn_unit, ])),'out': tf.Variable(tf.constant(0.1, shape=[output_size, ]))
}keep_prob = tf.placeholder(tf.float32)def lstm(batch_size):w_in=weights['in']b_in=biases['in']input=tf.reshape(X,[-1,input_size])  #需要将tensor转成2维进行计算，计算后的结果作为隐藏层的输入input_rnn=tf.matmul(input,w_in)+b_ininput_rnn=tf.reshape(input_rnn,[-1,time_step,rnn_unit])  #将tensor转成3维，作为lstm cell的输入#    cell1=tf.nn.rnn_cell.BasicLSTMCell(rnn_unit)
#    cell2=tf.nn.rnn_cell.BasicLSTMCell(rnn_unit)
#    cell=tf.nn.rnn_cell.MultiRNNCell(cells=[cell1,cell2])
#
#    init_state=cell.zero_state(batch_size,dtype=tf.float32)cell1=tf.nn.rnn_cell.BasicLSTMCell(rnn_unit)cell2=tf.nn.rnn_cell.BasicLSTMCell(rnn_unit)cell=tf.nn.rnn_cell.MultiRNNCell(cells=[cell1,cell2])lstm_cell=tf.nn.rnn.DropoutWrapper(cell,input_keep_prob=1.0,output_keep_prob=keep_prob)mlstm_cell = tf.nn.rnn.MultiRNNCell([lstm_cell]*lstm_layers,state_is_tuple=True)init_state= mlstm_cell.zero_state(batch_size,dtype=tf.float32)init_state=cell.zero_state(batch_size,dtype=tf.float32)with tf.variable_scope('scope', reuse=tf.AUTO_REUSE):#共享变量output_rnn,final_states=tf.nn.dynamic_rnn(cell, input_rnn,initial_state=init_state, dtype=tf.float32)  #output_rnn是记录lstm每个输出节点的结果，final_states是最后一个cell的结果output=tf.reshape(output_rnn,[-1,rnn_unit]) #作为输出层的输入w_out=weights['out']b_out=biases['out']pred=tf.matmul(output,w_out)+b_outreturn pred,final_states #train the LSTM model
def train_lstm():global batch_sizepred,_=lstm(batch_size)###损失函数###loss=tf.reduce_mean(tf.square(tf.reshape(pred,[-1,output_size])-tf.reshape(Y,[-1,output_size])))train_op=tf.train.AdamOptimizer(lr).minimize(loss)saver=tf.train.Saver(tf.global_variables())#只保留最后4次的模型参数,max_to_keep=4with tf.Session() as sess:sess.run(tf.global_variables_initializer())iter_time = 1000for i in range(iter_time):step=0start=0end=start+batch_sizewhile(end<len(train_x)):_,loss_=sess.run([train_op,loss],feed_dict={X:train_x[start:end],Y:train_y[start:end]})start+=batch_sizeend=start+batch_size#每10步保存一次参数if step%1000==0:print(i,step,loss_)if (step + 1) == len(train_x):saver.save(sess,folder,global_step=step,global_iter=i)step+=1#predict
def prediction(test_x,test_y):pred,_=lstm(1)saver=tf.train.Saver(tf.global_variables())with tf.Session() as sess:###参数恢复，调用已经训练好的模型###saver.restore(sess, folder) test_predict=[]for step in range(len(test_x)):prob=sess.run(pred,feed_dict={X:[test_x[step]]})predict=prob.reshape(-1,output_size)test_predict.extend(predict)test_y=scaler2.inverse_transform(test_y)y_test=test_y.reshape(len(test_x),output_size).Ttest_predict=scaler2.inverse_transform(test_predict)predict_test=test_predict.Ttime=np.arange(0,len(test_predict),1)depth=np.arange(0,output_size,1)T,D=np.meshgrid(time,depth)plt.figure()ax1=plt.subplot(311)ax1.invert_yaxis()plt.pcolormesh(T,D,y_test,cmap='jet')ax2=plt.subplot(312)ax2.invert_yaxis()plt.pcolormesh(T,D,predict_test,cmap='jet')        ax3=plt.subplot(313)ax3.invert_yaxis()error=y_test-predict_testplt.pcolormesh(T,D,error,cmap='jet')plt.show()train_lstm()
prediction(test_x,test_y)