实战：使用Prophet做个简单预测

数据格式

将下面的输入表格转为以ds和y为名字的两列，ds是时间，y是值。没什么理由，Prophet框架就是这样的（😓）

time	abc (随便列名)
2010/1/1	10.1
…
2010/12/1	10.1
2011/1/1	10.1
…
2011/12/1	10.1
…

计算ACF

直接from statsmodels.graphics.tsaplots import plot_acf使用plot_acf函数

从图上来看好像情况不错，那就接着用Prophet进行预测了。

使用Prophet预测

m = Prophet(growth='logistic', yearly_seasonality=2, changepoint_prior_scale=2, seasonality_mode='additive')
出现了几个离群点，除此之外拟合看起来不错哦。

完整代码

import pandas as pd
import matplotlib.pyplot as plt
from fbprophet import Prophet
import numpy as np
from fbprophet.plot import plot_yearly

from matplotlib.font_manager import FontProperties
from statsmodels.graphics.tsaplots import plot_acf, plot_pacf


def predict(history_cases, Period):
    df = pd.DataFrame()
    ds = [k for k in history_cases]
    y = [history_cases[k] for k in history_cases]
    df["ds"] = ds
    df["y"] = y
    temp = np.array(y[2:])
    print(temp)
    plot_acf(temp)  # 查看auto coefficient function，判断是否可以有季节性
    tran_test_index = len(df) - 3  # 除最后三个点外的其他点用来做训练，最后三个点用来做预测的验证
    df_train = df[:tran_test_index]
    y_min = 0
    y_max = df["y"].max() + 3
    # seasonality_prior_scale =12,
    # growth='logistic' 使用logistic必须添加上限（添加底下的'cap'）
    # yearly_seasonality = 2 傅里叶级数为2
    # changepoint_prior_scale 调节拟合程度
    # seasonality_mode='multiplicative' or additive 乘性季节周期模式【预测的季节性在时间序列开始时太大而在结束时太小】
    m = Prophet(growth='logistic', yearly_seasonality=2, changepoint_prior_scale=2,
                seasonality_mode='additive')
    df_train['cap'] = y_max
    m.fit(df_train)
    future = m.make_future_dataframe(periods=Period, freq='MS')
    future['cap'] = y_max
    forecast = m.predict(future)

    predictions = list(forecast['yhat'][-Period:])
    # 确保预测结果处于某个区间内
    for i in range(len(predictions)):
        if predictions[i] > y_max:
            predictions[i] = y_max
        if predictions[i] < y_min:
            predictions[i] = y_min
    # 绘制拟合图和预测点
    test_x = list(pd.to_datetime(df[tran_test_index:tran_test_index + Period]["ds"]))
    test_y = list(df[tran_test_index:tran_test_index + Period]["y"])
    fore_fig = m.plot(forecast)
    x_1 = forecast['ds'][-Period:]
    plt.scatter(test_x, test_y, color="red", s=150, label="truth")
    plt.scatter(x_1, predictions, color="green", marker='^', s=150, label="prediction")
    plt.legend()
    plt.show()
    print(predictions)
    return predictions


df = pd.read_csv("abc.csv", encoding="gbk")
df = df.interpolate(method='akima')
# df.fillna(0,inplace=True)
# df.dropna(axis=0,how='any')
# df = df.dropna(axis=0,how='any')

bi_history_cases = {}
location = 'abc'
Period = 2
for i in range(len(df)):
    bi_history_cases[str(df["time"][i])] = df[location][i]  # 添加进字典中

preds = predict(bi_history_cases, Period)

参考

自相关与偏自相关的简单介绍
Facebook 时间序列预测算法 Prophet 的研究（个人认为是知乎最牛Prophet回答）