AI驱动的供应链管理：需求预测实战指南

# 处理促销期间的异常峰值
def winsorize_promo(df, promo_col='is_promo', sales_col='sales'):
    promo_mask = df[promo_col] == 1
    q99 = df[promo_mask][sales_col].quantile(0.99)
    df.loc[promo_mask & (df[sales_col] > q99), sales_col] = q99
    return df

# 动态缺失值填充：相似SKU的同期均值
def dynamic_impute(df, sku_col='sku', date_col='date'):
    df['imputed'] = df.groupby([sku_col, df[date_col].dt.month])[sales_col].transform(
        lambda x: x.fillna(x.median())
    )
    return df

from tsfresh import extract_features
from tsfresh.feature_selection.relevance import calculate_relevance_table

# 自动提取300+时序特征
features = extract_features(
    df, 
    column_id='sku', 
    column_sort='date',
    default_fc_parameters=EfficientFCParameters()
)

# 基于假设检验筛选显著特征
relevance = calculate_relevance_table(features, df['demand'])
selected_features = relevance[relevance['p_value'] < 0.01]['feature'].tolist()

import lightgbm as lgb
from sklearn.metrics import mean_absolute_percentage_error as mape

# 构造滞后特征（过去7/14/30天销量）
def create_lags(df, lags=[7,14,30]):
    for lag in lags:
        df[f'sales_lag_{lag}'] = df.groupby('sku')['sales'].shift(lag)
    return df

# 自定义损失函数：加权MAPE（近期误差权重更高）
def weighted_mape(preds, train_data):
    labels = train_data.get_label()
    weights = np.power(0.9, np.arange(len(labels))[::-1])  # 指数衰减权重
    return 'weighted_mape', mape(labels, preds, sample_weight=weights), False

# 训练
train_data = lgb.Dataset(X_train, label=y_train)
model = lgb.train(
    params={
        'objective': 'regression_l1',
        'learning_rate': 0.03,
        'max_depth': 8,
        'num_leaves': 64,
        'feature_fraction': 0.8,
        'bagging_fraction': 0.8,
        'bagging_freq': 1
    },
    train_set=train_data,
    feval=weighted_mape,
    num_boost_round=1000,
    early_stopping_rounds=100
)

from pytorch_forecasting import TemporalFusionTransformer, Baseline

# 定义时间序列数据集
dataset = TimeSeriesDataSet(
    data=df,
    time_idx="time_idx",
    target="demand",
    group_ids=["sku", "warehouse"],
    min_encoder_length=30,
    max_encoder_length=60,
    min_prediction_length=7,
    max_prediction_length=14,
    static_categoricals=["category", "region"],
    time_varying_known_reals=["discount", "temperature"],
    time_varying_unknown_reals=["demand"],
    add_relative_time_idx=True,
    add_target_scales=True,
    add_encoder_length=True,
)

# 训练
tft = TemporalFusionTransformer.from_dataset(
    dataset,
    learning_rate=0.01,
    hidden_size=32,
    attention_head_size=4,
    dropout=0.1,
    loss=QuantileLoss(),
)
trainer.fit(
    tft,
    train_dataloaders=train_dataloader,
    val_dataloaders=val_dataloader,
)

# 注意力可视化：识别关键驱动因子
interpretation = tft.interpret_output(batch, reduction="sum")
tft.plot_interpretation(interpretation)

import shap

# 计算全局特征重要性
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)

# 发现：当折扣>30%且温度<5℃时，促销反而抑制需求（shap值<-0.3）
shap.dependence_plot(
    "discount", 
    shap_values, 
    X_test,
    interaction_index="temperature"
)

from kafka import KafkaConsumer
import faust

# 实时温度数据接入
app = faust.App('feature-pipeline', broker='kafka://localhost:9092')
temperature_topic = app.topic('weather', value_type=float)

@app.agent(temperature_topic)
async def process_temperature(temperatures):
    async for temp in temperatures:
        redis_client.set(f"temp_{datetime.now().date()}", temp)

import onnxmltools
onnx_model = onnxmltools.convert_lightgbm(model)
onnxmltools.utils.save_model(onnx_model, 'demand_predict.onnx')

from scipy.stats import ks_2samp

# 检测温度特征漂移
_, p_value = ks_2samp(
    reference_data['temperature'], 
    current_data['temperature']
)
if p_value < 0.05:
    trigger_model_retrain()