Code For Better 谷歌开发者之声——使用TensorFlow的时间序列预测

关注我的公众号YueTan进行交流探讨

一、 前言

对未来的预测能够帮助企业更好的把握当下。因此，时间序列任务广泛应用于交通、气象、金融、零售等行业。

本文介绍如何使用TensorFlow进行深度学习的时间序列预测，主要依托我即将发布的一个时序包东流TFTS (TensorFlow Time Series) 。TFTS是一个时间序列的开源工具，采用TensorFlow框架，支持多种深度学习SOTA模型。中文名“东流”，源自辛弃疾“青山遮不住，毕竟东流去。江晚正愁余，山深闻鹧鸪”。

1.1 开源地址

https://github.com/LongxingTan/Time-series-prediction

1.2 取得成绩

我曾使用该基于tensorflow的框架参加过一些数据竞赛，取得过一些成绩。

• Bert模型 获得KDD CUP2022百度风机功率预测第3名
• Seq2seq模型 获得阿里天池-AI earth人工智能气象挑战赛第4名

接下来，我们介绍面对一个新任务时，如何采用TensorFlow获得一个好的结果

二、应用案例

2.1 准备环境

• 可以选择docker安装，参照tensorflow文档
• 也可以参照tfts中的requirements

pip install -r requirements
1

2.2 准备数据

我们这里采用kaggle中的数据集，下载地址

我们首先对数据进行探索性分析，根据需求了解数据的情况。

2.3 特征工程

虽然，深度学习具有学习交叉特征的能力，但显式的特征工程仍然是不可获取的。

def add_features(df):
    df['cross']= df['u_in'] * df['u_out']
    df['cross2']= df['time_step'] * df['u_out']
    df['area'] = df['time_step'] * df['u_in']
    df['area'] = df.groupby('breath_id')['area'].cumsum()
    df['time_step_cumsum'] = df.groupby(['breath_id'])['time_step'].cumsum()
    df['u_in_cumsum'] = (df['u_in']).groupby(df['breath_id']).cumsum()
    print("Step-1...Completed")
    
    df['u_in_lag1'] = df.groupby('breath_id')['u_in'].shift(1)
    df['u_out_lag1'] = df.groupby('breath_id')['u_out'].shift(1)
    df['u_in_lag_back1'] = df.groupby('breath_id')['u_in'].shift(-1)
    df['u_out_lag_back1'] = df.groupby('breath_id')['u_out'].shift(-1)
    df['u_in_lag2'] = df.groupby('breath_id')['u_in'].shift(2)
    df['u_out_lag2'] = df.groupby('breath_id')['u_out'].shift(2)
    df['u_in_lag_back2'] = df.groupby('breath_id')['u_in'].shift(-2)
    df['u_out_lag_back2'] = df.groupby('breath_id')['u_out'].shift(-2)
    df['u_in_lag3'] = df.groupby('breath_id')['u_in'].shift(3)
    df['u_out_lag3'] = df.groupby('breath_id')['u_out'].shift(3)
    df['u_in_lag_back3'] = df.groupby('breath_id')['u_in'].shift(-3)
    df['u_out_lag_back3'] = df.groupby('breath_id')['u_out'].shift(-3)
    df['u_in_lag4'] = df.groupby('breath_id')['u_in'].shift(4)
    df['u_out_lag4'] = df.groupby('breath_id')['u_out'].shift(4)
    df['u_in_lag_back4'] = df.groupby('breath_id')['u_in'].shift(-4)
    df['u_out_lag_back4'] = df.groupby('breath_id')['u_out'].shift(-4)
    df = df.fillna(0)
    print("Step-2...Completed")
    
    df['breath_id__u_in__max'] = df.groupby(['breath_id'])['u_in'].transform('max')
    df['breath_id__u_in__mean'] = df.groupby(['breath_id'])['u_in'].transform('mean')
    df['breath_id__u_in__diffmax'] = df.groupby(['breath_id'])['u_in'].transform('max') - df['u_in']
    df['breath_id__u_in__diffmean'] = df.groupby(['breath_id'])['u_in'].transform('mean') - df['u_in']

    print("Step-3...Completed")
    
    df['u_in_diff1'] = df['u_in'] - df['u_in_lag1']
    df['u_out_diff1'] = df['u_out'] - df['u_out_lag1']
    df['u_in_diff2'] = df['u_in'] - df['u_in_lag2']
    df['u_out_diff2'] = df['u_out'] - df['u_out_lag2']
    df['u_in_diff3'] = df['u_in'] - df['u_in_lag3']
    df['u_out_diff3'] = df['u_out'] - df['u_out_lag3']
    df['u_in_diff4'] = df['u_in'] - df['u_in_lag4']
    df['u_out_diff4'] = df['u_out'] - df['u_out_lag4']
    print("Step-4...Completed")    
    return df

train = add_features(train)
test = add_features(test)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48

2.4 模型

tfts在tensorflow的基础上可以快速搭建不同时序模型。只需要传入对应模型名字即可

def build_model(use_model, train_sequence_length, predict_sequence_length=288, target_aggs=1, short_feature_nums=10, long_feature_nums=1, custom_model_params=None):
    inputs = (
        Input([1]),
        Input([train_sequence_length, short_feature_nums]),  # raw feature numbers
        Input([train_sequence_length+predict_sequence_length, long_feature_nums])  # 长版
        )
    teacher_inputs = Input([predict_sequence_length//target_aggs, 1])

    ts_inputs = KDD(train_sequence_length, predict_sequence_length)(inputs)
    outputs = build_tfts_model(use_model=use_model, predict_sequence_length=predict_sequence_length//target_aggs, custom_model_params=custom_model_params)(ts_inputs, teacher_inputs)
    model = tf.keras.Model(inputs={'inputs':inputs, 'teacher': teacher_inputs}, outputs=outputs)
    return model
1
2
3
4
5
6
7
8
9
10
11
12

目前已经支持的模型包括，点击可查看具体tensorflow实现

• RNN
• Seq2seq
• TCN
• wavenet
• bert
• transoformer

2.5 训练与验证

线下验证是衡量模型水平的重要依据，对于时序任务多采取time split的方式。

def run_train(cfg, args):    
    train_data_reader, train_data_loader, valid_data_reader, valid_data_loader, valid_df = build_data(cfg, args) 
    print(len(train_data_reader), len(valid_data_reader), len(valid_df))  

    build_model_fn = functools.partial(build_model, cfg.use_model, train_sequence_length=cfg.train_sequence_length, predict_sequence_length=cfg.predict_sequence_length, target_aggs=cfg.target_aggs, short_feature_nums=len(cfg.feature_column_short), long_feature_nums=len(cfg.feature_column_long), custom_model_params=cfg.custom_model_params)

    optimizer = tf.keras.optimizers.Adam(learning_rate = cfg.fit_params['learning_rate'])
    loss_fn = custom_loss  # tf.keras.losses.MeanSquaredError() 

    trainer = KerasTrainer(build_model_fn, loss_fn=loss_fn, optimizer=optimizer, strategy=None)
    trainer.train(train_data_loader, valid_dataset=valid_data_loader, **cfg.fit_params)
    valid_pred = trainer.predict(valid_data_loader, batch_size=cfg.fit_params['batch_size'] * 2)  # 3746 * 288    
    trainer.save_model(model_dir=cfg.model_dir + '/checkpoints/{}_day'.format(cfg.use_model), checkpoint_dir=cfg.checkpoint_dir+'/nn_day_{}.h5'.format(cfg.use_model))
    trainer.plot()
    
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

2.6 部署

利用训练过程中生成的二进制模型，可以采用tf-serving或Flask快速部署。

三、未来展望

做时序demo时，可以直接采用pip安装的方式进行，而做比赛的话，由于对精度有更高要求，模型本身可调节地方较多，建议修改源代码并从源码import的方式，例如取得成绩里的seq2seq和bert模型都是采用代码import的方式。

除了时间序列的多步预测任务之外，未来考虑结构不变的情况下，支持更多时序任务，欢迎支持

• 区间预测
• 异常检测
• 分类