ARIMA模型(英语:Autoregressive Integrated Moving Average model),差分整合移动平均自回归模型,又称整合移动平均自回归模型(移动也可称作滑动),是时间序列预测分析方法之一。ARIMA(p,d,q)中,AR是“自回归”,p为自回归项数;MA为“滑动平均”,q为滑动平均项数,d为使之成为平稳序列所做的差分次数(阶数)。“差分”一词虽未出现在ARIMA的英文名称中,却是关键步骤。
statsmodels.tsa.arima_model包中有ARIMA集成好的模型,我们只需要输入p,d,q即可。
数据为纽约市的交通进出情况(一个txt进,一个txt出),然后已知一年365天*24小时的数据,想用ARIMA来预测,并计算MAE和RMSE来评估预测准确性。
我的data_prepare()函数是用来拼接两个文件夹里的数据的,一般只要读取一个文件夹的数据即可,返回是一个大矩阵。
其中取了数据集的前66%做训练集,后34%做测试集,与预测的结果做对比。
res()函数是输入测试矩阵和预测矩阵,计算MAE和RMSE来评估预测准确性的。
main中写的三个数组是打算遍历p,d,q找到最优值的,但是我电脑跑的太慢了,最后直接取了0,1,0。
import warnings
import pandas as pd
import numpy as np
from sklearn.metrics import mean_squared_error
from sklearn.metrics import mean_absolute_error
from statsmodels.tsa.arima_model import ARIMA
def data_prepare():
matrix = []
file1 = "tensor_year_hour_lease.txt"
file2 = "tensor_year_hour_return.txt"
f1 = open(file1, "r")
f2 = open(file2, "r")
matrix1 = []
lines1 = f1.readlines()
for line in lines1:
arr = line.split(",")
arr = np.array(arr, dtype=int)
matrix1.append(arr)
f1.close()
matrix2 = []
lines2 = f2.readlines()
for line in lines2:
arr = line.split(",")
arr = np.array(arr, dtype=int)
matrix2.append(arr)
f2.close()
matrix = np.hstack((matrix1, matrix2)) # 拼接成功 输出(8760*188)
return matrix
def evaluate_arima_model(X, arima_order):
# 数据集的前66%作为训练集,后34%作为测试集
train_size = int(len(X) * 0.66)
# print("train_size",train_size)
train, test = X[0:train_size], X[train_size:]
history = [x for x in train]
# make predictions
predictions = list()
for t in range(len(test)):
model = ARIMA(history, order=arima_order)
model_fit = model.fit(disp=0)
yhat = model_fit.forecast()[0]
predictions.append(yhat)
history.append(test[t])
predictions = np.array(predictions)
# print(predictions.shape)
return predictions
def res(test, predictions):
mae = mean_absolute_error(test, predictions)
mse = mean_squared_error(test, predictions)
rmse = mse ** 0.5
return mae, rmse
def evaluate_models(p_values, d_values, q_values):
matrix = data_prepare()
train_size = int(len(matrix) * 0.66)
test = matrix[train_size:]
best_mae, best_rmse, best_cfg = float("inf"), float("inf"), None
for p in p_values:
for d in d_values:
for q in q_values:
pre = np.zeros((len(matrix)-train_size, 0))
for i in range(0, 188): # 188是列数
dataset = matrix[:, i]
dataset = dataset.astype('float32')
order = (p, d, q)
predictions = evaluate_arima_model(dataset, order)
pre = np.hstack((pre, predictions)) # 每一列做一次预测,然后拼接成矩阵
print("p, d, q, i:", p, d, q, i)
# print(pre.shape)
# print(pre)
mae, rmse = res(test, pre)
if mae < best_mae:
best_mae, best_rmse, best_cfg = mae, rmse, order
print('ARIMA%s MAE=%.3f RMSE=%.3f' % (order, mae, rmse))
print('Best ARIMA%s MAE=%.3f RMSE=%.3f' % (best_cfg, best_mae, best_rmse))
if __name__ == '__main__':
p_values = [0, 1, 2, 4, 6, 8, 10]
d_values = range(0, 3)
q_values = range(1, 3)
warnings.filterwarnings("ignore")
evaluate_models(p_values, d_values, q_values)