SKlearn MLP Regressor预测值持续为零问题求助

一、MLP预测值全零的修复方案 #

1. 核心问题定位 #

代码存在三个关键问题导致预测值为零：

• 时间特征处理错误：直接将datetime对象的numpy数组作为输入特征，MLP无法有效处理该类型输入，必须转换为数值型时间差。
• 目标变量未做尺度归一化：仅标准化输入特征而忽略目标变量，MLP对尺度敏感，容易收敛到全局均值或零值。
• MLP默认参数未适配回归任务：默认迭代次数不足、网络结构简单，模型未充分收敛。

2. 修改后的完整代码 #

import json
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.neural_network import MLPRegressor
from sklearn.preprocessing import RobustScaler
from sklearn.impute import SimpleImputer

# 加载训练数据
with open('data.json', 'r') as file:
    data = json.load(file)
df = pd.DataFrame(data)

# 正确转换时间特征：计算与训练集最早时间的秒数差
df['periodFrom'] = pd.to_datetime(df['periodFrom'])
min_time = df['periodFrom'].min()
df['periodFrom_numeric'] = (df['periodFrom'] - min_time).dt.total_seconds()

# 缺失值处理
imputer = SimpleImputer(strategy='mean')
df['value'] = imputer.fit_transform(df[['value']])

# 拆分特征与目标
X_train = df['periodFrom_numeric'].values.reshape(-1, 1)
y_train = df['value'].values

# 加载新数据
with open('new_data_json.json', 'r') as file:
    new_data = json.load(file)
new_df = pd.DataFrame(new_data)

# 新数据时间特征转换：使用训练集的最早时间作为基准
new_df['periodFrom'] = pd.to_datetime(new_df['periodFrom'])
new_df['periodFrom_numeric'] = (new_df['periodFrom'] - min_time).dt.total_seconds()
new_df['value'] = imputer.transform(new_df[['value']])
X_new = new_df['periodFrom_numeric'].values.reshape(-1, 1)

# 标准化输入特征与目标变量
scaler_X = RobustScaler()
X_train_scaled = scaler_X.fit_transform(X_train)
X_new_scaled = scaler_X.transform(X_new)

scaler_y = RobustScaler()
y_train_scaled = scaler_y.fit_transform(y_train.reshape(-1, 1)).ravel()

# 构建适配回归任务的MLP模型
model = MLPRegressor(
    hidden_layer_sizes=(64, 32),
    activation='relu',
    solver='adam',
    max_iter=1000,
    random_state=42,
    early_stopping=True,
    validation_fraction=0.1
)
model.fit(X_train_scaled, y_train_scaled)

# 预测并逆变换回原始尺度
predictions_scaled = model.predict(X_new_scaled).reshape(-1, 1)
predictions = scaler_y.inverse_transform(predictions_scaled)

# 可视化
plt.figure(figsize=(10, 6))
plt.scatter(df['periodFrom'], df['value'], label='训练集实际值', marker='o')
plt.scatter(new_df['periodFrom'], predictions, label='MLP预测值', marker='o', color="green")
plt.scatter(new_df['periodFrom'], new_df['value'], label='测试集实际值', marker="o", color="red")
plt.xlabel('时间周期')
plt.ylabel('流量值')
plt.title('实际值与预测值对比')
plt.legend()
plt.grid(True)
plt.xticks(rotation=45)
plt.tight_layout()
plt.show()

二、替代模型推荐 #

ARIMA残差过大说明数据可能存在非线性趋势或强季节性，以下模型更适配你的能源流量时间序列：

1. 梯度提升树（XGBoost/LightGBM） #

树模型对非线性模式、异常值鲁棒性强，无需数据平稳性假设，可直接处理时间特征：

import lightgbm as lgb

# 训练模型
model_lgb = lgb.LGBMRegressor(n_estimators=500, learning_rate=0.05, random_state=42)
model_lgb.fit(X_train, y_train)

# 预测
predictions_lgb = model_lgb.predict(X_new)

2. Prophet（Facebook开源工具） #

专门针对时间序列设计，自动捕捉趋势、年/周/日季节性，配置简单：

from prophet import Prophet

# 转换为Prophet要求的格式
prophet_train = df.rename(columns={'periodFrom': 'ds', 'value': 'y'})[['ds', 'y']]
prophet_test = new_df.rename(columns={'periodFrom': 'ds'})[['ds']]

# 训练与预测
model_prophet = Prophet(yearly_seasonality=True, weekly_seasonality=True, daily_seasonality=True)
model_prophet.fit(prophet_train)
forecast = model_prophet.predict(prophet_test)
predictions_prophet = forecast['yhat'].values

3. LSTM（循环神经网络） #

适合捕捉时间序列的长期依赖关系，对周期性流量数据效果优异：

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense

# 转换为序列输入格式（示例：用前3个时间步预测当前值）
def create_sequences(data, seq_length=3):
    X, y = [], []
    for i in range(len(data)-seq_length):
        X.append(data[i:i+seq_length])
        y.append(data[i+seq_length])
    return np.array(X), np.array(y)

# 准备序列数据
seq_X_train, seq_y_train = create_sequences(df['value'].values)
seq_X_train = seq_X_train.reshape(seq_X_train.shape[0], seq_X_train.shape[1], 1)

# 构建LSTM模型
model_lstm = Sequential()
model_lstm.add(LSTM(50, activation='relu', input_shape=(3, 1)))
model_lstm.add(Dense(1))
model_lstm.compile(optimizer='adam', loss='mse')
model_lstm.fit(seq_X_train, seq_y_train, epochs=50, batch_size=32)

内容的提问来源于stack exchange，提问作者Boyan