时间序列预测 — 指数平滑、因素分解、机器学习、深度学习、数据增强

置顶条件漫步

已于 2022-07-26 18:07:35 修改

阅读量1.7k

点赞数 4

分类专栏：时序预测文章标签：时序预测深度学习

于 2020-11-09 10:53:15 首次发布

本文链接：https://blog.csdn.net/chenhepg/article/details/109571042

版权

时序预测专栏收录该内容

15 篇文章 5 订阅

订阅专栏

@创建于：2020.11.09
@修改于：2022.05.18

1、经典方法

1.1 指数平滑

1.2 因素分解

1.2.1 朴素方法

1.2.2 ARIMA

1.2.3 Prophet

2、机器学习

3、深度学习

3.1 RNN/LSTM/GRU

3.2 Attention (Transformer)

3.3 DeepAR

CSDN：时间序列预测方法之 DeepAR

3.4 WaveNet

model = keras.models.Sequential()
model.add(keras.layers.InputLayer(input_shape=[60, 1]))
for rate in (1, 2, 4, 8) * 2:
    model.add(keras.layers.Conv1D(filters=20, kernel_size=2, padding="causal",
                                  activation="relu", dilation_rate=rate))
model.add(keras.layers.Conv1D(filters=10, kernel_size=1))
model.add(Flatten())
model.add(Dense(1, activation='sigmoid'))
model.compile(loss="mse", optimizer="adam")
model.summary()

3.5 Informer时序

3.6 Autoformer时序

from hfai.datasets import LTSF

dataset = LTSF(data_name='ETTh1', split='train', seq_len=96, label_len=48, pred_len=24, features='S', target='OT', timeenc=0, freq='h')
loader = dataset.loader(batch_size=64, num_workers=4)

from hfai.models import Autoformer
import hfreduce.torch as hfr

model = Autoformer(enc_in=1, dec_in=1, c_out=1, seq_len=96, label_len=48, out_len=24, e_layers=2, d_layers=1, dropout=0.05, embed='timeF').cuda()
loss = Loss(model)
optimizer = Optimizer(model)

reducer = hfr.AsyncReduceFloat(registry_host, port, proc_rank, procs_per_node, node_rank, node_cnt, model)

# each epoch
for x, y, x_mark, y_mark in loader:
    reducer.zero_grad() 
    
    model(x, y, x_mark, y_mark)
    loss.backward()
    
    reducer.synchronize()