R语言Prophet实现多维度分组批量时间序列预测方案咨询
多维度分组Prophet批量预测修改方案
前置处理(可选)
你可以先过滤y为0的记录,避免零值对模型效果产生负面影响:
df_clean <- df %>% filter(y != 0)
核心代码修改
仅需要调整nest的分组参数,将三个业务维度都设为分组键,每个唯一的(establishment, category, channel)组合会自动生成独立数据子集、对应训练一个Prophet模型:
兼容新旧版tidyr写法
d1 <- df_clean %>% nest(-establishment, -category, -channel) %>% mutate(m = map(data, prophet)) %>% # 按需调整period参数,预测2022年全年设为365即可 mutate(future = map(m, make_future_dataframe, period = 365)) %>% mutate(forecast = map2(m, future, predict)) d <- d1 %>% unnest(forecast) %>% # 按需保留预测字段,可追加yhat_lower、yhat_upper等置信区间字段 select(ds, establishment, category, channel, yhat)
新版tidyr(>=1.0.0)推荐标准写法
d1 <- df_clean %>% nest(data = -c(establishment, category, channel)) %>% mutate(m = map(data, prophet)) %>% mutate(future = map(m, make_future_dataframe, period = 365)) %>% mutate(forecast = map2(m, future, predict)) d <- d1 %>% unnest(forecast) %>% select(ds, establishment, category, channel, yhat)
效率优化建议
针对1000+模型的训练场景,可使用furrr包替换purrr的map函数实现并行训练,大幅降低运行耗时:
# 加载并行包 library(furrr) # 开启多线程,预留1核给系统 plan(multisession, workers = availableCores() - 1) d1 <- df_clean %>% nest(data = -c(establishment, category, channel)) %>% mutate(m = future_map(data, prophet)) %>% mutate(future = future_map(m, make_future_dataframe, period = 365)) %>% mutate(forecast = future_map2(m, future, predict))
内容的提问来源于stack exchange,提问作者Ed Cunningham
相关产品推荐
相关产品推荐

