在 scikit learn 中使用 svr 进行时间序列预测

Question

我有按日期索引的每日温度数据集，我需要使用 scikit-learn 中的 [SVR][1] 预测未来温度。

我坚持选择训练的 X 和 Y 以及测试的 X 放。例如，如果我想在时间 t 预测 Y，那么我需要 训练集包含t-1, t-2, ..., t-N 处的X 和Y，其中N 是用于预测Y 处t 的先前天数。

我该怎么做？

就是这样。

df=daily_temp1
# define function for create N lags
def create_lags(df, N):
    for i in range(N):
        df['datetime' + str(i+1)] = df.datetime.shift(i+1)
        df['dewpoint' + str(i+1)] = df.dewpoint.shift(i+1)
        df['humidity' + str(i+1)] = df.humidity.shift(i+1)
        df['pressure' + str(i+1)] = df.pressure.shift(i+1)
        df['temperature' + str(i+1)] = df.temperature.shift(i+1)
    df['vism' + str(i+1)] = df.vism.shift(i+1)
    df['wind_direcd' + str(i+1)] = df.wind_direcd.shift(i+1)
    df['wind_speed' + str(i+1)] = df.wind_speed.shift(i+1)
    df['wind_direct' + str(i+1)] = df.wind_direct.shift(i+1)

    return df

# create 10 lags
df = create_lags(df,10)


# the first 10 days will have missing values. can't use them.
df = df.dropna()

# create X and y
y = df['temperature']
X = df.iloc[:, 9:]

# Train on 70% of the data
train_idx = int(len(df) * .7)

# create train and test data
X_train, y_train, X_test, y_test = X[:train_idx], y[:train_idx], X[train_idx:], y[train_idx:]


# fit and predict
clf = SVR()
clf.fit(X_train, y_train)

clf.predict(X_test)

Answer 1

这是一个将特征矩阵 X 构建为简单的 lag1 - lagN 的解决方案，其中 lag1 是前几天的温度，而 lagN 是 N 天前的温度。

# create fake temperature
df = pd.DataFrame({'temp':np.random.rand(500)})

# define function for create N lags
def create_lags(df, N):
    for i in range(N):
        df['Lag' + str(i+1)] = df.temp.shift(i+1)
    return df

# create 10 lags
df = create_lags(df,10)

# the first 10 days will have missing values. can't use them.
df = df.dropna()

# create X and y
y = df.temp.values
X = df.iloc[:, 1:].values

# Train on 70% of the data
train_idx = int(len(df) * .7)

# create train and test data
X_train, y_train, X_test, y_test = X[:train_idx], y[:train_idx], X[train_idx:], y[:train_idx]

# fit and predict
clf = SVR()
clf.fit(X_train, y_train)

clf.predict(X_test)