用Prophet在Python中进行时间序列预测

使用fbprophet分析世界疫情感染人数

import pandas as pd
from fbprophet import Prophet
pred = pd.read_csv("../kaggle3/covid-19-all.csv")
pred.head()

	Country/Region	Province/State	Latitude	Longitude	Confirmed	Recovered	Deaths	Date
0	China	Anhui	31.8257	117.2264	1.0	NaN	NaN	2020-01-22
1	China	Beijing	40.1824	116.4142	14.0	NaN	NaN	2020-01-22
2	China	Chongqing	30.0572	107.8740	6.0	NaN	NaN	2020-01-22
3	China	Fujian	26.0789	117.9874	1.0	NaN	NaN	2020-01-22
4	China	Gansu	37.8099	101.0583	NaN	NaN	NaN	2020-01-22

pred = pred.fillna(0)
predgrp = pred.groupby("Date")[["Confirmed","Recovered","Deaths"]].sum().reset_index()
predgrp.head()

	Date	Confirmed	Recovered	Deaths
0	2020-01-22	555.0	28.0	17.0
1	2020-01-23	653.0	30.0	18.0
2	2020-01-24	941.0	36.0	26.0
3	2020-01-25	1438.0	39.0	42.0
4	2020-01-26	2118.0	52.0	56.0

观察前面五行，发现单日新增感染人数只有几百，再观察最近的感染人数，达到了恐怖的18万人之多！

predgrp.tail()

	Date	Confirmed	Recovered	Deaths
87	2020-04-18	2317759.0	592319.0	159510.0
88	2020-04-19	2401379.0	623903.0	165044.0
89	2020-04-20	2472259.0	645738.0	169986.0
90	2020-04-21	2549294.0	679819.0	176583.0
91	2020-04-22	2623415.0	709694.0	183027.0

pred_cnfrm = predgrp.loc[:,["Date","Confirmed"]]
pred_cnfrm.shape

(92, 2)

pr_data = pred_cnfrm
pr_data.columns = ['ds','y']
pr_data.head()

m=Prophet()
m.fit(pr_data)
future=m.make_future_dataframe(periods=15)   #预测15天
forecast=m.predict(future)
forecast.head().T

	0	1	2	3	4
ds	2020-01-22 00:00:00	2020-01-23 00:00:00	2020-01-24 00:00:00	2020-01-25 00:00:00	2020-01-26 00:00:00
trend	-1231.05	-218.877	793.296	1805.47	3092.11
trend_lower	-1231.05	-218.877	793.296	1805.47	3092.11
trend_upper	-1231.05	-218.877	793.296	1805.47	3092.11
yhat_lower	-11158.3	-8236.24	-3166.07	247.023	-982.845
yhat_upper	1602.52	5214.46	10051.4	12530.2	11597.8
additive_terms	-3974.05	-571.137	2528.97	4295.44	2357.41
additive_terms_lower	-3974.05	-571.137	2528.97	4295.44	2357.41
additive_terms_upper	-3974.05	-571.137	2528.97	4295.44	2357.41
multiplicative_terms	0	0	0	0	0
multiplicative_terms_lower	0	0	0	0	0
multiplicative_terms_upper	0	0	0	0	0
weekly	-3974.05	-571.137	2528.97	4295.44	2357.41
weekly_lower	-3974.05	-571.137	2528.97	4295.44	2357.41
weekly_upper	-3974.05	-571.137	2528.97	4295.44	2357.41
yhat	-5205.09	-790.014	3322.27	6100.91	5449.52

y_ped = forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']]
y_ped.head()

	ds	yhat	yhat_lower	yhat_upper
0	2020-01-22	-5205.094924	-11158.349533	1602.523589
1	2020-01-23	-790.014418	-8236.237859	5214.464592
2	2020-01-24	3322.270436	-3166.065919	10051.441647
3	2020-01-25	6100.908463	247.022750	12530.151876
4	2020-01-26	5449.521809	-982.844950	11597.844552

m.plot(forecast,xlabel='Date',ylabel='Confirmed Count', uncertainty=True)

这里是区间预测，但是由于数据比较大，所以看到点线近似重合了，所以认为误差可以被接受。

如果想查看预测的成分分析，可以使用 Prophet.plot_components 方法。
默认情况下，将展示趋势、时间序列的年度季节性和周季节性。 如果之前包含了节假日，也会展示出来。

m.plot_components(forecast)

误差分析：主要看绝对值误差，因为均方误差一般针对回归问题。

用sklearn库的内置函数计算误差

from sklearn.metrics import mean_squared_error # 均方误差
from sklearn.metrics import mean_absolute_error # 平方绝对误差

y = y_ped["yhat"].loc[0:91]

观察前几个数据的预测效果，发现差异比较大。

(y-pred_cnfrm["Confirmed"]).head()

0   -5760.094924
1   -1443.014418
2    2381.270436
3    4662.908463
4    3331.521809
dtype: float64

MSE（Mean Squared Error）均方误差

mean_squared_error(pred_cnfrm["Confirmed"],y)

24838490.129674848

MAE （Mean absolute Error）平均绝对误差

mean_absolute_error(pred_cnfrm["Confirmed"],y)

3423.4048112033311

利用numpy计算误差

MSE（Mean Squared Error）均方误差

import numpy as np
mse_test=np.sum((y-pred_cnfrm["Confirmed"])**2)/len(y) #跟数学公式一样的
mse_test

24838490.129674848

RMSE（Root Mean Squared Error）均方根误差

rmse_test=mse_test ** 0.5
rmse_test

4983.8228429263863

MAE （Mean absolute Error）平均绝对误差

mae_test=np.sum(abs(y-pred_cnfrm["Confirmed"]))/len(y) #跟数学公式一样的
mae_test

3423.4048112033311