R中ts()函数start参数原理及小时时间序列创建疑问

Let's tackle your two questions one by one, with concrete fixes for your code.

1. How does the start parameter in ts() work? #

The start argument in ts() takes a vector c(year, period), where period is defined entirely relative to the frequency you set:

• frequency represents the number of observations per full cycle (e.g., 12 for monthly data, 365 for daily, 8760 for hourly data across a non-leap year)
• The period value marks the exact position of your first observation within that cycle.

For example:

• If frequency=12 (monthly data), start=c(2017, 10) means the 10th observation of 2017—aka October.
• If frequency=8760 (hourly data for a non-leap year), start=c(2017, 6768) means the 6768th hour of 2017.

Your mistake was passing the day of year (282) as the period when frequency=8760. Since 282 hours into 2017 equals ~11.75 days, that’s why you ended up with 2017-01-12 instead of your intended start date.

2. Can you create an hourly time series with ts() for use with tslm()? #

Absolutely! You just need to correctly calculate the start period and set the right frequency. Here's how to fix your code:

Step 1: Calculate the correct starting hour position in the year #

Your first observation is 2017-10-09 00:00:00, which is the 282nd day of 2017. Convert this to its hour position in the year:

start_day_of_year <- as.numeric(format(df$interval[1], "%j"))
start_hour_in_year <- start_day_of_year * 24 + as.numeric(format(df$interval[1], "%H"))
# This gives 282*24 + 0 = 6768

Step 2: Create the ts object with correct parameters #

Use frequency=8760 (total hours in a non-leap year; use 8784 if your data spans a leap year):

df.ts <- ts(
  data = df$total,
  start = c(2017, start_hour_in_year),
  frequency = 24 * 365  # 8760 hours per non-leap year
)

Step 3: Verify the start date #

Check the starting date again with date_decimal() (from the lubridate package):

library(lubridate)
head(date_decimal(index(df.ts)), 1)
# Should return "2017-10-09 UTC" (with 00:00:00 depending on your locale)

Alternative: Focus on daily hourly seasonality #

If you want to model daily hourly patterns instead of annual cycles, set frequency=24 and define start using a decimal date for the first day:

start_decimal_date <- decimal_date(df$interval[1])
df.ts_daily_season <- ts(
  data = df$total,
  start = c(start_decimal_date, 0),
  frequency = 24
)

This structure works perfectly with tslm() and makes it easier to analyze daily trends.

内容的提问来源于stack exchange，提问作者Elena