Bayesian model leap year error?

[spencer-tassone]

I am trying to run a Bayesian metabolism model using a 15-minute time series collected over 600 days from May 2023 to January 2025. When I attempt to run the model with the complete time series, I encounter the following error after a while (I'm not quite sure how to interpret it):

Error in conquerList(results, naming = data.frame(year = data2[!duplicated(ids), : length(u1) == s1 * s2 is not TRUE

However, if I run the model excluding December 31, 2024, it runs successfully. Why can't the model handle the inclusion of December 31, 2024?

2024 was a leap year...

library(tidyverse)
library(dataRetrieval)
library(LakeMetabolizer)
library(streamMetabolizer) # to get irr estimates
library(forecastML) # for making time series evenly spaced

# 1. Download time series ----
wq_data <- readNWISuv(
  siteNumbers = "01670144", # Red House
  parameterCd = c("00010", "00300"), # 00010 = Wtemp (C), 00300 = DO (mgL)
  tz = "America/New_York"
) |>
  renameNWISColumns() |>
  mutate(
    k.gas = 0.55,
    z.mix = 1.8,
    do.sat = LakeMetabolizer::o2.at.sat.base(Wtemp_Inst, 960),
    solar.time = streamMetabolizer::calc_solar_time(dateTime, -77.9),
    irr = streamMetabolizer::calc_light(solar.time, 38.1, -77.9)
  ) |>
  rename(
    datetime = dateTime,
    wtr = Wtemp_Inst,
    do.obs = DO_Inst
  ) |>
  select(datetime, do.obs, do.sat, k.gas, z.mix, irr, wtr)

# 2. Make time series evenly spaced ----
wq_data <- forecastML::fill_gaps(
  data = wq_data,
  date_col = 1,
  frequency = "15 min"
)

# 3. Remove incomplete days at beginning and end of time series ----
wq_data <- wq_data |> 
  mutate(date = date(datetime)) |> 
  filter(date >= as.Date("2023-05-19") & date <= as.Date("2025-01-13")) |> 
  select(!date)

# 4. Interpolate missing data ----
wq_data <- wq_data |> 
  arrange(datetime) |> 
  mutate(do.obs = zoo::na.approx(do.obs, na.rm = FALSE),
         do.sat = zoo::na.approx(do.sat, na.rm = FALSE),
         k.gas = zoo::na.approx(k.gas, na.rm = FALSE),
         z.mix = zoo::na.approx(z.mix, na.rm = FALSE),
         irr = zoo::na.approx(irr, na.rm = FALSE),
         wtr = zoo::na.approx(wtr, na.rm = FALSE))

# 5. Run bayesian metabolism model (full time series; results in error) ----
metab <- LakeMetabolizer::metab(wq_data, "bayesian", lake.lat = 38.1)
# Error in conquerList(results, naming = data.frame(year = data2[!duplicated(ids),  : 
# length(u1) == s1 * s2 is not TRUE

# 6. Run subset bayesian metabolism model (excludes 2024-12-31; runs without error) ----
subset_dat <- wq_data |> 
  mutate(date = date(datetime)) |> 
  filter(!date == as.Date("2024-12-31")) |> 
  select(!date)

subset_metab <- LakeMetabolizer::metab(subset_dat, "bayesian", lake.lat = 38.1)

[spencer-tassone]

An alternative approach that appears to work for my situation is to subtract two years from each datetime observation, resulting in all dates falling within the range of 2021-2023 (non-leap years). In this workaround, I removed all data corresponding to the date 2024-02-29 before subtracting two years from each datetime observation.

[hdugan]

The problem is certainly a legacy of old code that these days would be written differently to hopefully prevent this error.
My suggestion would be running each year independently. I think if the data are subset by year and then run, you won't get the error.

### Split data by year
wq_list <- split(wq_data, year(wq_data$datetime))
### Apply metab() to each year's data
metab_list <- purrr::map(wq_list, ~{
  tryCatch({
    LakeMetabolizer::metab(.x, "bookkeep", lake.lat = 38.1)
  }, error = function(e) {
    message("Error in year ", unique(.x$year), ": ", e$message)
    return(NULL)
  })
})
### Bind list 
metab_results = bind_rows(metab_list)