NeuralProphet之七：NeuralProphet + Optuna

时间序列论文: NeuralProphet: Explainable Forecasting at Scale
NeuralProphet之一：安装与使用
NeuralProphet之二：季节性(Seasonality)
NeuralProphet之三：回归(Regressors)
NeuralProphet之四：事件(Events)
NeuralProphet之五：多时序预测模型
NeuralProphet之六：多元时间序列预测
NeuralProphet之七：NeuralProphet + Optuna
NeuralProphet之八：NeuralProphet部署
NeuralProphet官方示例一：建筑物用电量预测(Building load forecasting)
NeuralProphet官方示例二：日照辐射强度预测(Forecasting hourly solar irradiance)

NeuralProphet之七：NeuralProphet + Optuna

NeuralProphet
optuna

数据来自真实园区数据，
训练集为2022年8月01日到2022年8月14日园区的耗电量，每隔5分钟采集一次；
测试集为2022年8月15日到2022年8月21日园区的耗电量，每隔5分钟采集一次；
要求：预测接下来1H的耗电量。

代码和数据地址：https://github.com/shanglianlm0525/TimeSeries

导入库

from copy import deepcopy
import optuna
import pandas as pd
import matplotlib.pyplot as plt
from neuralprophet import NeuralProphet, set_log_level
set_log_level("ERROR")
1
2
3
4
5
6

导入数据

df_train = pd.read_csv("shenghuoqu0815.csv")
df_test = pd.read_csv("shenghuoqu0821.csv")
print(df_train.info())
print(df_test.info())
1
2
3
4

1 Baseline

n_lags = 12
n_forecasts=12*1
# n_lags = n_lags,n_forecasts= n_forecasts,
# changepoints_range=0.85, n_changepoints=20,
m = NeuralProphet(n_lags = n_lags,n_forecasts= n_forecasts,
                  yearly_seasonality=False, weekly_seasonality=False, daily_seasonality=True,
                  normalize='minmax')

metrics = m.fit(df_train, freq='5min')
forecast_train = m.predict(df_train)
forecast_test = m.predict(df_test)

fig = m.plot(forecast_train)
fig = m.plot(forecast_test)
fig_param = m.plot_parameters()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

训练集测试结果：

测试集测试结果：

参数可视化：

2 加入Optuna

定义Optuna参数的搜索范围

default_para = dict(n_lags = n_lags,n_forecasts= n_forecasts,changepoints_range=0.8,n_changepoints=5, trend_reg=0.1,normalize='minmax',learning_rate =1)

param_types = dict(changepoints_range='float',n_changepoints='int',trend_reg='float',learning_rate='float',month_order = 'int',week_order ='int')
bounds = {'changepoints_range': [0.6,0.8,0.9],
          'n_changepoints': [4, 8],
          'trend_reg': [0.001, 1],
          'learning_rate': [0.001, 1],
          'day_order': [1, 7],
         }
1
2
3
4
5
6
7
8
9

定义搜索空间

def create_nph(**para):
    temp_para = deepcopy(para)
    day_order = temp_para.pop('day_order')
    m = NeuralProphet(**temp_para)
    m = m.add_seasonality('my_day', 24, day_order)
    return m

def nph_warper(trial,ts):
    params = {}
    params['changepoints_range'] = trial.suggest_categorical('changepoints_range', bounds['changepoints_range'])
    params['n_changepoints'] = trial.suggest_int('n_changepoints', bounds['n_changepoints'][0], bounds['n_changepoints'][1])
    params['trend_reg'] = trial.suggest_loguniform('trend_reg', bounds['trend_reg'][0], bounds['trend_reg'][1])
    params['learning_rate'] = trial.suggest_loguniform('learning_rate', bounds['learning_rate'][0], bounds['learning_rate'][1])
    params['day_order'] = trial.suggest_int('day_order', bounds['day_order'][0], bounds['day_order'][1])
    temp_para = deepcopy(default_para)
    temp_para.update(params)
    METRICS = ['SmoothL1Loss', 'MAE', 'RMSE']
    metrics_test = pd.DataFrame(columns=METRICS)
    m = create_nph(**temp_para)
    folds = m.crossvalidation_split_df(ts, freq="H", k=5, fold_pct=0.2, fold_overlap_pct=0.5)
    for df_train, df_test in folds:
        m = create_nph(**temp_para)
        train = m.fit(df_train)
        test = m.test(df=df_test)
        metrics_test = metrics_test.append(test[METRICS].iloc[-1])
    out = metrics_test['MAE'].mean()
    return out
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

定义objective 函数（目标函数）

def objective(trial):
    ts = df_train.copy() #select column associated with region
    return nph_warper(trial,ts)
1
2
3

Optuna 调优

study = optuna.create_study(direction='minimize',study_name='nph_tuning', load_if_exists=True, storage="sqlite:///nph.db")
study.optimize(objective, n_trials=5)
print(study.best_trial)
print(study.best_params)
1
2
3
4

FrozenTrial(number=3, values=[157.7605354309082], datetime_start=datetime.datetime(2022, 8, 22, 18, 48, 27, 661116), datetime_complete=datetime.datetime(2022, 8, 22, 18, 49, 17, 188990), params={‘changepoints_range’: 0.6, ‘day_order’: 1, ‘learning_rate’: 0.005463461412824059, ‘n_changepoints’: 7, ‘trend_reg’: 0.006287378033976726}, distributions={‘changepoints_range’: CategoricalDistribution(choices=(0.6, 0.8, 0.9)), ‘day_order’: IntUniformDistribution(high=7, low=1, step=1), ‘learning_rate’: LogUniformDistribution(high=1.0, low=0.001), ‘n_changepoints’: IntUniformDistribution(high=8, low=4, step=1), ‘trend_reg’: LogUniformDistribution(high=1.0, low=0.001)}, user_attrs={}, system_attrs={}, intermediate_values={}, trial_id=15, state=TrialState.COMPLETE, value=None)
{‘changepoints_range’: 0.6, ‘day_order’: 1, ‘learning_rate’: 0.005463461412824059, ‘n_changepoints’: 7, ‘trend_reg’: 0.006287378033976726}

根据最优参数构建NeuralProphet

best_para = deepcopy(default_para)
best_para.update(study.best_params)

m = create_nph(**best_para)
1
2
3
4

训练和预测

metrics = m.fit(df_train, freq='5min')
forecast_train = m.predict(df_train, decompose=False)
forecast_test = m.predict(df_test, decompose=False)

fig = m.plot(forecast_train)
fig = m.plot(forecast_test)
fig_param = m.plot_parameters()
1
2
3
4
5
6
7

明显可以看出，Optuna找到的参数在测试集上效果更好。
训练集测试结果：

测试集测试结果：

参数可视化：