Published
Edited
Sep 7, 2022
Forecasting stream temperature
Predicting temperature exceedance
How likely is it that stream temperature will exceed 75F?
Ideas:
show how forecasts change with different # of ensemble models, provide explanation of sources of uncertainty
show how forecasts change as issue time approaches actual date. how accurate are we 7 vs 3 vs 1 day out?
the forecast data is limited to a single site (1450), with issue time from 6/27-7/07. my thought is to show predictions for 3 days at each issue time to see how they improve through time.
Overview of data and possible dimensions
lots of ensembles w/ different predictions, and final mean prediction
For a given date, making prediction 8 days out up through 0 days out
accuracy of predictions across lead times
Accuracy of predictions w/i single lead time regardless of date
E.g. day 1 prediction accuracy
First chart
timeseries + 90% CI - adjust # of ensembles and which lead times shown, and see how forecast differs
opacity, so that see overlap
live calculating mean of all shown ensembles
actual observations
Data: what's in here (forecast_1450 - 10 issue dates of forecasts (for all lead times), with all ensembles
NOTE: forecast_1450 doesn't include the CIs
function get_issue_dates(in_data) {
let date_dict = {}
for (var l=0; l <in_data.length; l++) {
let raw_date = in_data[l].issue_time;
let formatTime = d3.timeFormat("%Y-%m-%d");
let formatted_date = formatTime(raw_date)
date_dict[formatTime(raw_date)] = [] //{} date_dict[formatTime(raw_date)]
// date_dict[formatTime(raw_date)]['raw_date'] = raw_date//in_data[l].issue_time
// date_dict[raw_date] = []
// date_dict[raw_date]['raw_date'] = in_data[i].issue_time//in_data[l].issue_time
}
return(date_dict)
}
function get_line_chart_data(chart_1_data, ensembles) {
// get array of ensembles
let ensembleArray = makeEnsembleArray(ensembles)
// Use first ensemble to define issue dates (since same for all ensembles)
let ensembleData_1 = filterDataToEnsemble(chart_1_data, 1)
let issue_dates_dict = get_issue_dates(ensembleData_1)
// iterate over issue dates to populate dictionary w/
// data for each ensemble on each issue date
// let all_data_array = []
let issue_date;
for (issue_date in issue_dates_dict) {
// let issue_date_raw = issue_dates_dict[issue_date]['raw_date']
for (var m = 0; m<ensembleArray.length; m++) {
var ensembleData = filterDataToEnsemble(chart_1_data, m+1)
issue_dates_dict[issue_date][m] = filterEnsembleDataToIssueDate(ensembleData, m+1, issue_date)
// all_data_array.push(filterEnsembleDataToIssueDate(ensembleData, m+1, issue_date))
}
}
return(issue_dates_dict)
// return(all_data_array)
}
// ensembleArray.forEach(function(ensemble_num) {
// let ensembleData = filterDataToEnsemble(dataIn_1, ensemble_num)
// let ensembleIssueDates = {}
// for (var i=0; i< ensembleData.length; i++) {
// ensembleIssueDates[ensembleData[i].issue_time] = {}
// };
// for (var issue_date in ensembleIssueDates) {
// data[issue_date] = {}
// data[issue_date][ensemble_num] = filterEnsembleDataToIssueDate(ensembleData, issue_date)
// }
// data['issue dates'] = ensembleIssueDates
// // data[ensemble_num] = {};
// // data[ensemble_num]['issue_dates'] = ensembleIssueDates;
// // for (var issue_date in data[ensemble_num]['issue_dates']) {
// // data[ensemble_num]['issue_dates'][issue_date] = filterEnsembleDataToIssueDate(ensembleData, issue_date)
// // }
// })
Second chart
concept = focus on difference between observed and predicted, show change in CI over lead times
facet(?) diagram -- confidence intervals of the predictions and you can toggle how many days out you show predictions for - each lead time = new row
toggles: lead time
Data: 1 date, at all the different lead times, precalculate mean and CI for predictions
chart_2_data = forecasts.filter(d => formatDate(d.time) === filterDate)
function get_chart2_data(forecasts, filterDate) {
let single_day = []
single_day = forecasts.filter(d => formatDate(d.time) === filterDate)
single_day = tidy(single_day,
mutate({time: d => formatDate(d.time)}))
let summary = []
summary = tidy(
single_day,
groupBy(
['issue_time','time'],
summarize({mean_max_temp: mean('max_temp'),
min_max_temp: min('max_temp'),
max_max_temp: max('max_temp'),
n_ensembles: n()})
)
)
return(summary)
}
chart_2_data_summary = get_chart2_data(forecasts, filterDate)
forecasts[7].time === forecasts[15].time
new Date(forecasts[7].time) == new Date(forecasts[15].time)
formatDate = d3.timeFormat("%Y-%m-%d")
formatDate(forecasts[7].time)
formatDate(forecasts[7].time) === formatDate(forecasts[15].time)
filterDate = formatDate(forecasts[7].time)
forecasts
chart_2_data.length
Third chart
concept = highlight accuracy for a given lead time - each row = lead time
Timeseries bubble matrix, with dots scaled by size according to difference between observed and predicted, color light to dark where dark = small confidence interval
x-axis = date (time)
y-axis = lead time (day 7 at top) (time - issue time)
see this punchcard plothttps://observablehq.com/@observablehq/integration-test-flakiness
could later add toggle for site
Later can encode rmse, too
Precalculate radius and CI metric - range of CI across all dates and all lead times
Data: One site, max temp pred for X dates for all lead times
Data in
forecasts = FileAttachment("forecasts_1450.csv").csv({ typed: true })
Inputs.table(forecasts)
SummaryTable(forecasts)
Imports
d3 = require("d3@^6.1")
import { SummaryTable } from "@observablehq/summary-table"
Purpose-built for displays of data
Observable is your go-to platform for exploring data and creating expressive data visualizations. Use reactive JavaScript notebooks for prototyping and a collaborative canvas for visual data exploration and dashboard creation.
