深度学习：基于TensorFlow的双层BiDirection_Attention_LSTM的北京PM2.5数据集预测

基于TensorFlow的双层BiDirection_Attention_LSTM的北京PM2.5数据集预测

预测北京PM2.5数据集你会学到：

传统的线性模型难以解决多变量或多输入问题，而神经网络如LSTM则擅长于处理多个变量的问题，该特性使其有助于解决时间序列预测问题。
　　
　　在接下来的这篇博客中，你将学会如何利用深度学习框架TF搭建LSTM模型来处理多个变量的时间序列预测问题。
　　经过这个博客你会掌握：
　　1. 如何将原始数据转化为适合处理时序预测问题的数据格式；
　　2. 如何准备数据并搭建双层双向＋Attention的LSTM来处理时序预测问题；
　　3. 如何利用模型预测。

1.数据集内容

2.数据处理

利用Pandas处理数据集，PM2.5数据集在这里 ==》
[link](Beijing PM2.5 Data Set)

def PreData():
    df = pd.read_csv(config.preprocess_file_path)
    # 丢弃掉不需要的行
    df.drop('No', axis=1, inplace=True)
    # 丢弃含有NA的行
    df.dropna(axis=0, how='any', inplace=True`
    # 填充csv中的NAN为0.0
    df.fillna(axis=0, value=0.0, inplace=True)

    # index为False表示不要序列
    df.to_csv(config.process_file_path, index=False)

    new_pd = pd.read_csv(config.process_file_path)
    print(new_pd.info())
    print(new_pd.head(5))
    
    #准备训练和测试的批次为100的数据
    
def getTrain_Test():
    df = pd.read_csv(config.process_file_path)
    weather_data = df.iloc[0:40000, 5:12]
    test_data = df.iloc[40000:41700, 5:12]
    weather_pm = df.iloc[0:40000, 4:5]
    test_pm = df.iloc[40000:41700, 4:5]

    print(weather_data)
    print(type(weather_pm))
    print(len(weather_pm))

    i = 0
    # 当数据为测试集时
    if FLAGS.isPreData_isTrain_isPredict == 2:
        while i < len(test_data):
            # 当数据为训练集时

            x_train = test_data[i:i + config.batch_size].values
            y_label = test_pm[i:i + config.batch_size].values
            # print(type(x_train))   <class 'numpy.ndarray'>
            # print(x_train.shape)   (72, 7)
            string_encode = sp.LabelEncoder()
            to_1 = sp.MinMaxScaler(feature_range=(0, 1))
            x_train[:, 3] = string_encode.fit_transform(x_train[:, 3])
            x_train = to_1.fit_transform(x_train)
            y_label = to_1.fit_transform(y_label)
            x_train = np.reshape(x_train, [100, 7, 1])
            train_x_list.append(x_train.tolist())
            label_y_list.append(y_label.tolist())
            i += config.batch_size

    if FLAGS.isPreData_isTrain_isPredict == 1:
        while i < len(weather_pm):
            # 当数据为训练集时

            x_train = weather_data[i:i + config.batch_size].values
            y_label = weather_pm[i:i + config.batch_size].values
            # print(type(x_train))   <class 'numpy.ndarray'>
            # print(x_train.shape)   (72, 7)
            string_encode = sp.LabelEncoder()
            to_1 = sp.MinMaxScaler(feature_range=(0, 1))
            x_train[:, 3] = string_encode.fit_transform(x_train[:, 3])
            x_train = to_1.fit_transform(x_train)
            y_label = to_1.fit_transform(y_label)
            x_train = np.reshape(x_train, [100, 7, 1])
            train_x_list.append(x_train.tolist())
            label_y_list.append(y_label.tolist())
            i += config.batch_size

3.需要导入的包以及参数配置

import tensorflow as tf
import pandas as pd
import sklearn.preprocessing as sp
import numpy as np

FLAGS = tf.flags.FLAGS
tf.flags.DEFINE_integer('isPreData_isTrain_isPredict', 1, '0为PreData，1为Train，2为Predict')
tf.flags.DEFINE_integer('isAttention', 1, '0为关闭Attention机制，1为启用Attention机制')
tf.flags.DEFINE_integer('is_featureMap', 0, '0为每个特征都Attention权重，1为每个featureMap一个权重')


class Config():
    preprocess_file_path = './PRSA_data_2010.1.1-2014.12.31.csv'
    process_file_path = './PMcsv/bjpm.csv'
    batch_size = 100
    loop_time = [128, 128]
    input_train_dim = 7
    input_label_dim = 1
    lstm_units = 128


config = Config()
train_x_list = []
label_y_list = []

input_data = tf.placeholder(tf.float32, [1, config.batch_size, config.input_train_dim, 1])
input_label = tf.placeholder(tf.float32, [1, config.batch_size, config.input_label_dim])

4.下面上LSTM模型了

def Bi_lstm():
    global input_data
    # 双层LSTM，第二次LSTM将前向，后向的输出值拼接起来
    concat_tensor = input_data
    concat_tensor = tf.reshape(concat_tensor, [100, 7, 1])

    with tf.variable_scope('lstm1'):
        lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=config.lstm_units)
        lstm_fw_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell, input_keep_prob=0.95, output_keep_prob=0.95)
        lstm_bw_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell, input_keep_prob=0.95, output_keep_prob=0.95)
        # [100,7,1]===>[100,7,128]
        value, _ = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell, lstm_bw_cell, concat_tensor, dtype=tf.float32)
        # [[100,7,128],[100,7,128]]===>[100,7,256]
        concat_tensor = tf.concat(value, 2)
        # concat_tensor = tf.layers.batch_normalization(concat_tensor)
    with tf.variable_scope('lstm2'):
        lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=config.lstm_units)
        lstm_fw_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell, input_keep_prob=0.90, output_keep_prob=0.80)
        lstm_bw_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell, input_keep_prob